Multi-source information fusion fault diagnosis method and system combining equipment atlas and AI model

Abstract

The invention discloses a multi-source information fusion fault diagnosis method combining an equipment graph and an AI model. The method comprises the following steps: in an offline stage, preparing an industrial asset historical data set and constructing an equipment topology knowledge graph; training a double-machine learning diagnosis engine comprising a diagnosis home model and a health assessment model; and constructing an unstructured knowledge base based on the technical document. In the online diagnosis stage, the output of the diagnosis home model is monitored, and the diagnosis process is triggered when the state is abnormal; activating a diagnosis planner to query the equipment atlas to obtain an associated component set; calling a health assessment model to obtain a quantitative health index; health comparison is carried out through knowledge retrieval and logical reasoning; when evidence conflicts occur, iterative reasoning is carried out until a root cause is confirmed; generating a structured evidence file; and finally, an attribution analyzer is activated to output a professional diagnosis report. A system for implementing the method is also disclosed. According to the invention, accurate, automatic and interpretable diagnosis of faults is realized, and the operation and maintenance efficiency is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial intelligent operation and maintenance and fault diagnosis, and in particular to a multi-source information fusion fault diagnosis method and system combining device graph and AI model. BACKGROUND

[0002] In modern industry, especially in high-value asset-intensive fields such as aerospace, intelligent manufacturing, and energy and power, predictive and health management is a core supporting technology for ensuring the safe and reliable operation of equipment, reducing unplanned downtime, and optimizing maintenance strategies. With the popularization of industrial Internet of Things and sensor technology, data-driven fault diagnosis and prediction methods have become the mainstream research direction.

[0003] Currently, the closest prior art to the present application mainly includes the following: 1. Supervised learning-based fault diagnosis technology: This technology collects a large amount of sensor data of the equipment under various known fault modes, trains a machine learning or deep learning classification model. In actual operation, the model analyzes real-time sensor data snapshots to determine whether the equipment is currently malfunctioning and the type of fault.

[0004] 2. Unsupervised learning-based anomaly detection technology: In the absence of sufficient fault label data, this technology uses clustering algorithms or reconstruction models to model the data distribution of the equipment under normal operating conditions. When a new data point deviates significantly from the known "healthy" data pattern, the system will determine it as "abnormal". This technology can discover unknown or rare anomalies, but it is usually unable to directly point out the specific cause and location of the anomaly.

[0005] 3. Knowledge-based expert system: This technology attempts to encode the diagnostic knowledge of human experts into a fixed rule base through a series of "IF-THEN" rules. The system triggers the corresponding rules through a logical reasoning engine based on the input parameters to give a diagnostic conclusion.

[0006] 4. Information retrieval-based auxiliary diagnosis system: This technology digitizes and indexes unstructured documents such as equipment technical manuals and historical maintenance cases. Maintenance personnel can search for relevant information related to the current problem through keyword search to assist in decision-making. In recent years, some systems have begun to introduce retrieval augmentation generation (RAG) technology, using large language models (LLM) to provide more intelligent question-and-answer retrieval.

[0007] Although the above technologies have achieved some success in specific scenarios, they generally have the following deep defects and shortcomings when dealing with increasingly complex industrial systems: 1. Lack of holistic perspective on complex system failure analysis: Whether supervised or unsupervised learning models, they are essentially doing mathematical pattern matching on a set of isolated data. They do not understand the physical location and functional relationship of a component in a complex system. Therefore, when systemic problems occur, causing multiple component sensors to be abnormal, such models cannot distinguish which is the root cause and which is the downstream complication caused by the root cause. This leads to one-sided diagnosis focusing only on the local, and cannot conduct in-depth root cause tracing.

[0008] 2. Fragmentation of multi-source heterogeneous information: Existing technologies lack means to effectively integrate different sources of unstructured information. Multi-source different data, such as operation data, operation manual, past documents are often used to train different models, and there is a lack of effective integration method. Past combination methods often rely on human to integrate information, and lack a method to automatically form a logically rigorous, mutually corroborative diagnostic evidence chain.

[0009] 3. Lack of dynamic strategy planning ability: The diagnostic logic of existing systems is static and linear. Supervised learning and rule-based expert systems completely rely on pre-set rules and cannot cope with new situations outside the rule base; while the unsupervised learning paradigm based on no label often lacks interpretability, can only identify abnormal behavior, and is difficult to specify the fault type, and thus difficult to effectively correct the real fault and guide. SUMMARY

[0010] The present application provides an intelligent diagnosis method based on the overall perspective of a complex system, which integrates structured system knowledge, unstructured text knowledge and model quantitative analysis results, to solve the problem of context missing in traditional methods and achieve deeper root cause analysis.

[0011] The application provides a multi-source information fusion fault diagnosis method combining equipment graph and AI model, which has the following characteristics: an offline preparation stage and an online diagnosis stage; the offline preparation stage comprises the following steps: step 1, preparing an industrial asset historical data set and constructing an equipment topology knowledge graph; step 2, training a double-machine learning diagnosis engine comprising a “diagnosis home” model and a “health assessment” model according to the industrial asset historical data set; and step 3, constructing an unstructured knowledge base based on industrial asset technical documents; the online diagnosis stage comprises the following steps: step 4, monitoring the output of the “diagnosis home” model, triggering the diagnosis process when the state jumps from normal to abnormal, and obtaining a preliminary diagnosis result; step 5, activating an “diagnosis strategist” agent based on the preliminary diagnosis result, querying the equipment topology knowledge graph, and thus obtaining a set of associated components of the fault component; step 6, performing health assessment on the associated components by calling the “health assessment” model, and obtaining a quantitative health index; step 7, the “diagnosis strategist” agent receives the quantitative health index, compares it with the overall health status of the asset through knowledge retrieval, and performs logical reasoning; step 8, when there is a conflict in the evidence, the “diagnosis strategist” agent dynamically generates a revised diagnosis execution strategy and iteratively reasons until the root cause component is confirmed, and finally obtains a root cause reasoning conclusion, thereby realizing clear fault positioning; step 9, the “diagnosis strategist” agent packages the preliminary diagnosis result, the set of associated components and the root cause reasoning conclusion to generate a structured evidence file; step 10, activating a “cause analyzer” agent, and the “cause analyzer” agent converts the structured evidence file into a deep analysis report; and step 11, the “cause analyzer” agent outputs a professional diagnosis report according to the attribution rules by taking the deep analysis report as the final context, wherein the professional diagnosis report comprises analysis of the current root cause and operation suggestions.

[0012] In the multi-source information fusion fault diagnosis method combining equipment graph and AI model provided by the application, the industrial asset historical data set in step 1 can further comprise multivariate time series data, health state data labels and technical documents; the multivariate time series data X is a three-dimensional array , wherein N is the total number of asset units, is the maximum running time, and D is the running signal feature dimension of each unit; the health state data label is that each asset unit in the training set is labeled with its corresponding health state label at each time point t, and the label is a quantitative health degree index, residual life or qualitative fault mode classification; and the technical documents are various unstructured or semi-structured text materials related to the asset, which are used to construct the unstructured knowledge base. The construction method is as follows: First, based on the publicly available design drawings, P&ID diagrams, or equipment lists of industrial assets, extract their core physical components. V is the node set of the knowledge graph; then, based on the physical flow, spatial relationships, or logical dependencies between components, a directed edge set E is defined, where each directed edge... Representation Component It is a component The upstream of or directly affected by it, and the specific relationship is recorded in natural language.

[0013] The multi-source information fusion fault diagnosis method combining device mapping and AI models provided by this invention may also have the following feature: In step 2, dual machine learning diagnostic engines are deployed in the background API service, including a "diagnostic expert" model and a "health assessment" model. The "diagnostic expert" model... This is a multi-class classification model used for preliminary fault identification and localization, which uses a certain time t or time window as a reference. Single-device D-dimensional feature vector or all devices 3D feature vector As input, the output is the health status of a specific asset or the overall asset at that moment, and the posterior probability distribution of belonging to K+1 categories. The model comprises K+1 categories, including K known abnormal states and 1 normal state. (Health Assessment Model) This refers to one or more time series prediction models that provide a quantitative and probabilistic assessment of the future health status of equipment components, using a D-dimensional feature sequence of a component over W past time steps. As input, its output It is any one or a combination of two of the continuous health indicator prediction values ​​and the probability distribution vector of future failures.

[0014] The multi-source information fusion fault diagnosis method combining equipment maps and AI models provided by this invention may also have the following feature: In step 4, the process of obtaining preliminary diagnostic results is as follows: the monitoring module transmits the feature vectors collected in real time at a preset frequency... Transmitted to the "Diagnostician" model Perform a state check when the model's highest probability output first changes from the "normal state" category to an "abnormal state" category. When the system detects a fault, it formally triggers a full-process diagnostic process. The trigger signal is defined as a tuple. In the formula, n represents the asset ID and t represents the trigger time. This represents the preliminary diagnostic results. In step 5, the process of obtaining the associated component set of the faulty component is as follows: After the "Diagnostic Strategist" is activated, it receives a trigger signal. It comes from Extract the faulty component name And use it as a parameter, combined with specific fault results Call the device map The query tool performs a graph traversal algorithm and returns the results related to the faulty node. Directly related upstream component set and downstream component sets .

[0015] The multi-source information fusion fault diagnosis method combining device maps and AI models provided by this invention may also have the following feature: In step 6, the process of obtaining the quantified health index is as follows: the "diagnosis strategist" agent targets the key related component set returned in the previous step. and Call the "health assessment" model Conduct independent health assessments for each upstream component. The system extracts its relevant feature sequences and inputs them into the system. To obtain an independent quantitative health index In step 7, the logical reasoning process is as follows: the "diagnostic strategist" large language model agent receives the quantified health index returned in step 6. In conjunction with the external multimodal database, it compares the data with the overall health status of the assets through knowledge retrieval and performs logical reasoning. Its core task is to comprehensively evaluate the heterogeneous evidence collected in the preceding steps. The evidence includes at least qualitative labels for preliminary diagnosis, structured relationships in the knowledge graph, and quantitative values ​​for health assessment.

[0016] The multi-source information fusion fault diagnosis method combining device maps and AI models provided by this invention may also have the following feature: In step 8, when the system identifies logical contradictions or insufficient confidence levels between different sources of evidence, the "diagnostic strategist" agent identifies such abnormal states and uses its reasoning ability to dynamically generate a corrected diagnostic execution strategy to obtain new evidence from the system. The corrected strategy may involve calling additional diagnostic tools for further data collection or marking the conflict in the final evidence file. The aforementioned dynamic diagnostic execution strategy is an iterative process, specifically: after each round of evidence collection, the "diagnostic strategist" large language model agent evaluates the current evidence file containing qualitative, quantitative, and structured evidence; the iterative reasoning process continues until the "diagnostic strategist" agent determines that the evidence meets preset conditions and stops, confirming the root cause component.

[0017] The multi-source information fusion fault diagnosis method combining device maps and AI models provided by this invention may also have the following feature: In step 9, the "diagnosis strategy" large model intelligent agent will use the preliminary diagnosis results obtained in step 2... The list of associated components obtained in step 5 The root cause reasoning conclusions, as confirmed in step 8, are packaged into a structured JSON evidence file. In step 10, the process of generating the in-depth analysis report is as follows: The "attribution analyzer" agent transforms the structured evidence archive into an in-depth analysis report. This process strictly follows the following two-stage analysis and generation flow: Stage 1, preliminary evidence extraction and factual statement: The "attribution analyzer" agent first receives the structured evidence archive. As internal evidence, a preliminary fact report is generated by paraphrasing and summarizing the objective facts contained therein. During this process, the agent does not introduce external knowledge, make subjective inferences, or establish causal relationships not explicitly stated in the evidence. Phase two, enhanced knowledge retrieval and comprehensive analysis: After completing the fact statement, the "attribution analyzer" agent first dynamically generates an enhanced query based on the content of the "preliminary fact report." Subsequently, the large model agent uses this query to initiate a vector retrieval from the unstructured knowledge base to obtain relevant external knowledge.

[0018] The multi-source information fusion fault diagnosis method combining device maps and AI models provided by this invention may also have the following features: In step 11, the attribution rules are: each statement must be clearly attributed to its source; when the evidence is insufficient, it must be stated truthfully without fabricating answers; all causal relationships must be based on the content clearly stated in internal evidence or external knowledge without creating new causal chains.

[0019] This invention provides a multi-source information fusion fault diagnosis system combining equipment topology and AI models, characterized by the following features: an offline preparation module for performing the following operations: constructing an equipment topology knowledge graph; training a dual machine learning diagnostic engine containing a "diagnostician" model and a "health assessment" model based on historical industrial asset datasets; constructing an unstructured knowledge base based on industrial asset technical documents; and an online diagnosis module connected to the offline preparation module, comprising the following parts: a model-driven fault triggering unit that monitors the output of the "diagnostician" model and triggers the diagnostic process when the state changes from normal to abnormal to obtain preliminary diagnostic results; a "diagnostic strategist" agent activation and correlation analysis unit that, based on the preliminary diagnostic results, activates the "diagnostic strategist" agent to query the equipment topology knowledge graph to obtain the associated component set of the faulty component; and a cross-validation evidence collection unit that, for the associated component set, calls the "health assessment" model to perform a health assessment on the associated components to obtain a quantitative health index. The system comprises the following components: a cross-validation logical reasoning unit, where the "Diagnostic Strategist" agent receives a quantified health index, compares it with the overall health status of the assets through knowledge retrieval, and performs logical reasoning; an iterative reasoning and root cause confirmation unit, where, when evidence conflicts, the "Diagnostic Strategist" agent dynamically generates a revised diagnostic execution strategy and performs iterative reasoning until the root cause component is confirmed, ultimately arriving at a root cause reasoning conclusion, thus achieving clear fault location; an evidence aggregation and transfer unit, where the "Diagnostic Strategist" agent packages the preliminary diagnostic results, related component sets, and root cause reasoning conclusions into a structured evidence archive; a multi-source information fusion unit, which activates the "Attribution Analyzer" agent, which transforms the structured evidence archive into a deep analysis report; and an attributable report generation unit, where the "Attribution Analyzer" agent uses the deep analysis report as the final context and outputs a professional diagnostic report according to attribution rules, including analysis of the current root cause and operational suggestions.

[0020] The present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the method of the present invention when executing the program. Attached Figure Description

[0021] Figure 1 This is a flowchart of the offline preparation stage in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram and equipment diagram of Embodiment 1 of the present invention; Figure 3 This is a flowchart of the multi-source information fusion fault diagnosis method combining equipment maps and AI models in Embodiment 1 of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0023] Example 1 This implementation addresses the challenges of existing data-driven fault diagnosis methods, such as the lack of a holistic perspective, fragmented multi-source information, and insufficient dynamic strategy planning capabilities. It provides a multi-source information fusion fault diagnosis method that combines equipment topology mapping with AI models. By constructing a dual machine learning engine comprising a "diagnostician" classification model and a "health assessment" prediction model, and combining equipment topology mapping with an unstructured text knowledge base, a dual-agent collaborative architecture consisting of a "diagnostic strategist" and an "attribution analyzer" is designed. This architecture intelligently fuses heterogeneous information such as numerical, structural, and textual data into a complete diagnostic evidence chain through a series of logical reasoning steps, including "association analysis" and "model-driven cross-validation." This enables in-depth root cause analysis of complex industrial assets, significantly improving the accuracy, reliability, and interpretability of the diagnosis, and providing precise and reliable references for industrial asset operation and maintenance decisions.

[0024] The method and system proposed in this invention will be described in detail below through specific implementation methods, using a publicly available industrial simulation dataset as an example.

[0025] The validation dataset used in this embodiment is the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset. This dataset simulates the complete process of a turbofan engine from healthy operation to failure under different operating conditions and failure modes. Each dataset contains multivariate time-series data for multiple engine units, including 21 sensor readings and 3 operating setpoints.

[0026] The multi-source information fusion fault diagnosis method combining device maps and AI models in this embodiment includes an offline preparation stage and an online diagnosis stage.

[0027] First, in the offline phase, based on historical datasets and technical documents of industrial assets, a device topology knowledge graph, a dual machine learning diagnostic engine including a "diagnostician" classification model and a "health assessment" prediction model, and an unstructured knowledge base are constructed. In the online diagnostic phase, this method defines a dual-agent collaborative workflow consisting of a "diagnostic strategist" and an "attribution analyzer": First, the monitoring module triggers a deep diagnostic process based on the real-time classification results of the "diagnostic expert" model, transitioning from "normal" to "abnormal," and obtains preliminary fault component localization. Next, the first agent, the "diagnostic strategist," is activated. It calls the device atlas for correlation analysis to obtain upstream and downstream information of the faulty component and continuously calls the "health assessment" model to perform model-driven cross-validation of related components. The "diagnostic strategist" then integrates information and traces root causes until the key abnormal component is located. All collected evidence is then packaged into a structured JSON file. Finally, the second agent, the "attribution analyzer," receives this evidence file and, using it as context, initiates an enhanced knowledge retrieval from the knowledge base. This intelligently fuses all heterogeneous information, including numerical, structural, and textual data, to generate a final, attributable diagnostic report containing a complete chain of evidence. The specific implementation is as follows: Figure 1 This is a flowchart of the offline preparation stage in Embodiment 1 of the present invention.

[0028] like Figure 1 As shown, before conducting online diagnostics, the construction of the system's three core capability modules must first be completed. The offline preparation phase in this embodiment includes the following steps: Step S1: Prepare the historical dataset of industrial assets and construct a knowledge graph of equipment topology. The historical dataset of industrial assets contains the following types of data: Multivariate time series data X: a three-dimensional array In the formula, N represents the total number of asset units. Where D is the maximum runtime, and D is the dimension of the operating signal characteristics of each unit; Health status data tags For each asset unit in the training set at each time point t, it is labeled with its corresponding health status label. This label can be a quantitative health index / remaining lifetime (RUL) or a qualitative failure mode classification.

[0029] Technical documentation: This includes various unstructured or semi-structured text documents related to the asset, such as equipment operation and maintenance manuals, historical failure case reports, and process flow diagrams. This data is used to build the RAG knowledge base, providing in-depth domain knowledge support for the agent's final reasoning.

[0030] This embodiment uses the C-MAPSS dataset as the "Historical Dataset of Industrial Assets". Specifically: Process and state data X: Use the 24 features (3 setpoints + 21 sensors) contained in the dataset as input features for the model.

[0031] Supervision information data : Quantitative health status indicators: For a run-to-failure dataset containing a single failure mode, these indicators are calculated for each engine in each cycle. The remaining number of flight cycles ( ), and obtain labels for training the "health assessment" model.

[0032] Qualitative fault category labeling: For run-to-failure datasets containing multiple fault modes (such as high-pressure compressor and fan), this is achieved by analyzing key sensors (such as...) associated with the high-pressure compressor and fan during the end-of-life phase (e.g., the last 10% of cycles) of each engine. and The system intelligently labels each engine with its root cause of failure ("high-pressure compressor degradation" or "fan degradation") based on the rate of degradation. Simultaneously, it extracts health data samples from the first 30% of the lifecycle of each engine unit in the single failure mode dataset and labels them as "normal." These labels are used to train the "diagnostician" model.

[0033] Technical documentation: Official C-MAPSS technical documentation is used as the source of the unstructured knowledge base.

[0034] Device topology knowledge graph The construction method is as follows: First, based on the publicly available design drawings, P&ID diagrams, or equipment lists of industrial assets, extract their core physical components. V is the node set of the knowledge graph; then, based on the physical flow, spatial relationships, or logical dependencies between components, a directed edge set E is defined, where each directed edge... Representation Component It is a component The upstream of the network or its direct influence is recorded in natural language. This network graph is constructed and stored using a network analysis library, providing structured knowledge for subsequent association analysis. The specific implementation method is as follows: Figure 2 This is a schematic diagram of Embodiment 1 of the present invention and its equipment diagram.

[0035] like Figure 2 As shown, the map is constructed and stored for loading by the backend API service.

[0036] In this embodiment, based on the engine structure diagram and airflow path diagram provided in the official C-MAPSS technical documentation, its core physical components are extracted as the node set of the knowledge graph. In this embodiment, Based on the physical path of the airflow, directed connections are established between nodes. (e.g., edge) This indicates that the low-pressure compressor is upstream of the high-pressure compressor.

[0037] Step S2: Based on the historical dataset of industrial assets, train a dual machine learning diagnostic engine comprising a "diagnostician" model and a "health assessment" model. The dual machine learning diagnostic engine is deployed in the backend API service and includes two models trained offline. "Diagnostician" model This is a multi-class classification model used for preliminary fault identification and localization, which uses a certain time t or time window as a reference. Single-device D-dimensional feature vector or all devices 3D feature vector As input, the output is the health status of a specific asset or the overall assets at that moment. Posterior probability distribution of each category (K known abnormal states and 1 normal state) .

[0038] "Health Assessment" Model A time series forecasting model that provides a quantitative and probabilistic assessment of the future health status of equipment, based on a D-dimensional feature sequence over the past W time steps. As input, its output It can be one of the following two forms or a combination thereof: (a) Continuous health indicator predictions: such as predicted remaining lifespan, or a health score between 0 and 1.

[0039] (b) Probability distribution vector of future failures: For example, output a vector ,in This represents the probability that the asset will experience a type i failure within a future time window.

[0040] In this embodiment, the training process of the dual machine learning diagnostic engines is as follows: Training the "diagnostic" model : To quickly and accurately identify the current system health status, this step trains a machine learning model to diagnose the health status of engine components at a given moment. The "qualitative fault category label" data prepared in step S1 (containing samples of three categories: "normal," "high-pressure compressor decline," and "fan decline") is used as the training set. Since this task is based on pattern recognition using a snapshot of data at the current moment, this embodiment employs an XGBoost multi-classification model. The model uses feature vectors from all feature dimensions... As input, the output is the posterior probability distribution of belonging to the three categories at that moment. After training, the "Diagnostician" model is saved and deployed as a component of the backend diagnostic engine for use in the online diagnostic phase.

[0041] Training the "health assessment" model : To quantitatively predict and assess the engine's health status, this step trains a machine learning model to diagnose the overall engine health status at a given moment. This step uses the operation-to-failure data (including remaining service life labels) prepared in step S1, which includes single-failure modes, as the training set. This embodiment employs a Long Short-Term Memory (LSTM) network model, based on past... One flight cycle dimensional feature sequence As input, the output is a prediction of the remaining useful life at the current moment. After training, the "health assessment" model is saved and deployed as a component of the backend diagnostic engine for use in the online diagnostic phase.

[0042] Step S3: Construct an unstructured knowledge base based on industrial asset technical documents.

[0043] This embodiment uses the Dify platform as the orchestration and hosting environment for the intelligent agent. The technical documents prepared in step S1 are uploaded to Dify, and the Dify platform automatically segments and vectorizes them, and builds a vector knowledge base that can be used by the intelligent agent for retrieval, enhancement, generation, and invocation.

[0044] In addition, the offline preparation phase in this embodiment also includes the encapsulation process of the background API service: A backend API service is built using the FastAPI framework, which is responsible for loading all assets (models, graphs, and data) generated in steps S1 through S3. It exposes three core API tools for the Dify agent to use: Status diagnostic tool: Receives device identifier and time point identifier, and calls... The model returns the current state.

[0045] Health assessment tool: Receives device identifier and time point identifier, and calls... The model returns a predicted quantitative health index (e.g., remaining lifespan value).

[0046] Device Knowledge Graph Query Tool: Receives device identifiers and queries the knowledge graph. Returns a structured list containing its upstream and downstream associated components.

[0047] Figure 3 This is a flowchart of the multi-source information fusion fault diagnosis method combining equipment maps and AI models in Embodiment 1 of the present invention.

[0048] like Figure 3 As shown, this embodiment uses a test data stream representing the complete lifecycle of a single engine, extracted from a dataset containing multiple fault modes, as the simulated data stream for online diagnostics. The online diagnostic phase includes the following steps: Step S4, Model-driven fault triggering: The monitoring module transmits the real-time collected feature vectors at a preset frequency. Transmitted to the "Diagnostician" model Perform a state check when the model's highest probability output first changes from the "normal state" category to an "abnormal state" category. When the system detects a fault, it formally triggers a full-process diagnostic process. The trigger signal is defined as a tuple. In the formula, n represents the asset ID and t represents the trigger time. This represents the preliminary diagnostic results.

[0049] In this embodiment, the monitoring program simulates data stream playback and calls the status diagnostic tool interface at each flight cycle t. In the initial phase ( ), The model continuously outputs "normal". When the simulation runs to... When the highest probability category returned by the interface changes from "normal" to "high-pressure compressor decline" for the first time, the system determines that a fault has occurred and formally triggers in-depth diagnosis.

[0050] At this point, the front-end application sends a trigger signal to the first agent on the Dify platform, the "Diagnostic Strategist". The request is sent via a variable in Dify: {engine_id: 1, cycle: 195, initial_diagnosis: 'HPC_Degradation'} Step S5, "Diagnostic Strategist" agent activation and correlation analysis: The system's first agent, "Diagnostic Strategist," is activated and receives a trigger signal. It comes from Extract the faulty component name And use it as a parameter, combined with specific fault results Call the device map A query tool. This tool performs a graph traversal algorithm and returns the results related to faulty nodes. Directly related upstream component set and downstream component sets .

[0051] In this embodiment, after the "Diagnostic Strategist" agent is activated, it is prompted to begin performing logical reasoning. It extracts preliminary diagnostic results from the input variables. After retrieval enhancement, knowledge related to the degradation of the high-pressure compressor is extracted and logical reasoning is performed. The "diagnostic strategist" agent, after reasoning, concludes that this component may not be the root cause of the engine problem, and thus issues the first tool call command to the backend API service: call_tool("get_related_components", component_name="HPC") Backend API query knowledge graph Returns a list of components associated with the high-pressure compressor node: {"component":"HPC", "upstream":["LPC"], "downstream":["Combustor"]} Step S6, evidence collection for cross-validation: The "diagnostic strategist" agent targets the set of key related components returned in the previous step. and Call the "health assessment" model Conduct independent health assessments for each upstream component. The system extracts its relevant feature sequences and inputs them into the system. To obtain an independent quantitative health index .

[0052] In this embodiment, the "Diagnostic Strategist" agent receives the graph query results from the previous step and learns that the low-pressure compressor is upstream of the high-pressure compressor. The "Diagnostic Strategist" agent begins to perform cross-validation, first assessing the health status of the upstream components. It issues a second tool call instruction to the backend API service: call_tool("predict_rul", engine_id=1, current_cycle=150) Backend API calls the "health assessment" model The current data window is evaluated. Data from the multi-failure mode dataset does indeed contain faults, and the model returns a low remaining lifetime value. {"engine_id":1, "predicted_rul":14.95} Based on the above analysis, the "diagnostic strategist" agent confirmed that the overall health of the system was low and initially identified the degradation of some equipment components as a possible second root cause.

[0053] Step S7, Logical reasoning for cross-validation: The "diagnostic strategist" large language model agent receives the quantified health index returned in step S6. It also combines the external multimodal database with knowledge retrieval to compare it with the overall health status of the assets and perform logical reasoning.

[0054] Unlike mechanical execution of pre-set fixed rules, the core task of the "diagnostic strategist" large language model agent is to comprehensively evaluate the heterogeneous evidence collected in the preceding steps. This evidence includes, but is not limited to, qualitative labels from preliminary diagnoses, structured relationships in knowledge graphs, and quantitative values ​​from health assessments. Through prompting engineering and managing the contents of external databases, the agent can synthesize the chain of evidence to perform root cause analysis of abnormal states and handle ambiguities, uncertainties, and conflicts between pieces of evidence.

[0055] In this embodiment, the "Diagnostic Strategist" agent reports "high-pressure compressor degradation" (from step S4), and the "Health Assessment" model reports a low remaining predicted lifetime of 14.95 (from step S6). The agent determines that all evidence is logically consistent and the preliminary diagnosis is credible. The "Diagnostic Strategist" agent reports a dynamic revision of the diagnostic strategy, concluding that the signal degradation of the high-pressure compressor may be a complication symptom transmitted from the degradation problem of the upstream low-pressure compressor, and pointing suspicion to the upstream low-pressure compressor component as the root cause.

[0056] In other possible embodiments, the initial diagnosis is high-pressure compressor degradation, but this is achieved by invoking a dataset trained on a subset of high-pressure compressor features. The model determines that the upstream low-pressure compressor components still have a high remaining predicted life. The "diagnostic strategist" agent then determines that the fault is caused by internal issues within the high-pressure compressor or by reverse transmission from downstream equipment, ruling out the possibility of upstream transmission.

[0057] Step S8, Iterative Reasoning and Root Cause Identification: When the system identifies logical contradictions or insufficient confidence levels between different sources of evidence, the agent can recognize such anomalies and, using its reasoning capabilities, dynamically generate a revised diagnostic execution strategy to extract new evidence from the system. The revised strategy includes either invoking additional diagnostic tools for further data collection or marking the conflict in the final evidence archive, thereby achieving a more robust and in-depth analysis than traditional fixed processes.

[0058] The dynamic diagnostic execution strategy is an iterative process. After each round of evidence collection, the "diagnostic strategy agent" of the large language model comprehensively evaluates the completeness, consistency, and clarity of the current evidence archive (including all qualitative, quantitative, and structured evidence). This iterative reasoning process continues, repeatedly calling tools and acquiring new evidence until the "diagnostic strategy agent" of the large language model determines that the currently collected evidence is sufficient to form a logical closed loop, the conflict has been resolved, and the root cause definition of the fault has reached a preset confidence threshold, thereby achieving clear fault localization.

[0059] In this embodiment, the dynamic diagnostic execution strategy enters its second iteration. In this iteration, the "diagnostic strategy agent" identifies the low-pressure compressor from the previous round as a new suspected fault component and again invokes the "device graph query tool" to obtain its upstream component, namely "fan 1". Next, the agent invokes the "health assessment tool" to perform an independent health status assessment for the "fan" component. In this embodiment, the assessment tool returns a high remaining predicted lifetime value.

[0060] At this point, the "Diagnostic Strategist" agent evaluates the complete chain of evidence currently collected: "Fan" (healthy) -> "Low-pressure compressor" (unhealthy) -> "High-pressure compressor" (unhealthy). Based on this chain, the "Diagnostic Strategist" agent utilizes its enhanced retrieval capabilities to query technical documents for knowledge of fault propagation and evolution patterns between "Fan 1" and the "Low-pressure compressor" (e.g., how a certain type of "Fan 1" problem evolves into a "Low-pressure compressor" failure after a period of time). It uses its knowledge reasoning capabilities to comprehensively analyze this newly collected heterogeneous evidence (i.e., fan health data, retrieved document knowledge, and the conclusion that the low-pressure compressor is unhealthy). In this embodiment, the quantitative assessment result of the "Fan" returned by the "Health Assessment Tool" is "healthy," and this data evidence refutes the hypothesis formed by the agent based on knowledge reasoning. The "Diagnostic Strategist" agent ultimately deduces that the source of the fault propagation has been found, determining that the low-pressure compressor degradation is the root cause of this diagnostic event. At this point, the chain of evidence has formed a logical closed loop, and the root cause of the fault is clearly defined. Therefore, the iteration process ends.

[0061] Step S9, Evidence Gathering and Transfer: The "Diagnostic Strategist" large model agent gathers all the original evidence collected in the above steps, including the preliminary diagnostic results from step S2. The list of associated components in step S5 The root cause reasoning conclusions finally confirmed in step S8 are packaged into a structured JSON evidence file. .

[0062] In this embodiment, the "diagnostic strategist" agent packages all the original evidence collected in the above steps into a structured JSON evidence archive. And use that as the final output. "engine_id": 1, "current_cycle": 150 "initial_diagnosis":"HPC_Degradation", "related_components": {"upstream": ["LPC"], "downstream": ["Combustor"]}, "system_health_assessment": {"predicted_rul": 14.95} } Step S10, Multi-source information fusion: The "attribution analyzer" large model agent will transfer the machine-readable data from the previous step. The evidence file is transformed into a fully attributable and reliable in-depth analysis report. This process strictly follows a two-stage analysis and generation workflow: Phase 1: Preliminary Evidence Extraction and Fact Statement: The "Attribution Analyzer" agent first receives... As internal evidence. At this stage, its task is only to [provide evidence to] [the relevant parties]. The objective facts contained herein are paraphrased and summarized. This step strictly prohibits the agent from introducing any external knowledge, making subjective inferences, or establishing causal relationships not explicitly stated in the evidence, thereby generating a "preliminary fact report" based purely on internal evidence.

[0063] Phase Two: Enhanced Knowledge Retrieval and Comprehensive Analysis: After completing the factual statement, the "attribution analyzer" agent enters the second phase. It first dynamically generates a semantically rich enhanced query based on the content of the "preliminary fact report." Subsequently, the large model agent uses this query to initiate a vector retrieval from the unstructured knowledge base to obtain relevant external knowledge, such as relevant chapters in technical manuals.

[0064] Step S11, Attributable report generation In the final step, the "attribution analyzer" large model agent combines the "preliminary fact report" (i.e., internal evidence) generated in stage one with the "external knowledge" retrieved in stage two as the final context. The agent then performs the final synthesis and generation according to a strict set of attribution rules: A. Rigorous Evidence-Based Approach: Every statement in the final report must be clearly attributed to its source; B. Fabrication is strictly prohibited: If the evidence is insufficient to support a part of the analysis (such as "maintenance recommendations"), the large model agent must state the facts truthfully, rather than fabricating answers; C. Strict Logic: Any causal relationship described in the report must be explicitly stated in "internal evidence" or "external knowledge," and the creation of new causal chains is not permitted.

[0065] Through these two-stage processes, the intelligent agent ultimately outputs a professional diagnostic report that is logically rigorous, has credible conclusions, and whose original evidence can be traced for every key piece of information. The report includes an analysis of the current root cause and operational suggestions.

[0066] Specifically, in this embodiment, the "attribution analyzer" large language model agent enters the final multi-source information fusion and report generation stage. In this stage, the agent uses the "external knowledge" retrieved in step S10 and the "internal evidence" transferred in step S9 together as the complete context for its reasoning. The agent performs the generation task according to a set of predefined strict attribution rule prompts: Every statement in the final report must be verifiable against '#1. Internal Evidence' (i.e., The evidence must be directly supported by the internal evidence archive or '#2. External Knowledge Base Retrieval' (i.e., the retrieved technical documents). Introducing any external information not provided in the context or subjective speculation is strictly prohibited. Its task is to 'synthesize' the heterogeneous information provided to build a coherent analysis. If the evidence in a section, even after incorporating external knowledge, remains weak or nonexistent, it must be stated truthfully that the information provided is insufficient to provide recommendations, not that recommendations are fabricated. Any causal chains described in the 'Causal Analysis' and 'Root Cause Conclusion' sections must be explicitly stated in '#2. External Knowledge Base Retrieval' (i.e., the referenced documents). The use of speculative language to create new causal relationships not mentioned in the documents is strictly prohibited. The presented logic must strictly mirror the logic of its source. 'External knowledge' can only be used to interpret 'internal evidence' and understand the root cause of 'this particular event'. General definitions or information from the knowledge base that are not directly related to the events described in the internal log must not be included in the report.

[0067] It intelligently integrates all heterogeneous information and ultimately outputs a professional diagnostic report. A key feature of this report is that every crucial conclusion explicitly cites its original source, thus achieving the ultimate goal of reliability and traceability in this embodiment.

[0068] Table 1 presents a quantitative comparison of the two methods in terms of "accuracy rate" and "expert consensus rate" of the generated reports. "Accuracy rate" assesses whether the report effectively addresses the diagnostic task and provides actionable modification suggestions; "expert consensus rate" assesses whether the report's analytical logic is completely consistent with expert-defined best diagnostic practices. The results show that the method proposed in this embodiment significantly outperforms the benchmark method in both metrics. Notably, the benchmark RAG method has a very low "expert consensus rate," indicating that while it can generate grammatically correct reports, its analytical logic and conclusions deviate significantly from expert-defined best practices and are more prone to logical fabrication. This invention, through a two-stage approach and strict attribution rules, increases the "expert consensus rate" to 90.00%, ensuring the report's logical rigor and high reliability.

[0069] Table 1. Quality Comparison of Different Report Generation Methods Auxiliary variable selection method Accuracy (%) Expert agreement rate (%) Benchmark RAG method 95.00 79.17 This example (two-stage attribution) 97.50 90.00 Example 2 This implementation provides a multi-source information fusion fault diagnosis system that combines equipment maps and AI models, including an offline preparation module and an online diagnosis module: The offline preparation module in this embodiment is used for the operations in steps S1 to S3 of embodiment 1; In this embodiment, the online diagnostic module is connected to the offline preparation module and includes the following parts: The model-driven fault triggering unit is used to perform the operation of step S4 in embodiment 1; The "Diagnostic Strategist" agent is initiated and associated analysis unit is used to perform the operation of step S5 in Example 1; The evidence collection unit for cross-validation is used to perform the operation of step S6 in Example 1; The cross-validation logic reasoning unit is used to perform the operation of step S7 in Example 1; The iterative reasoning and root cause identification unit is used to perform the operation of step S8 in Example 1; The evidence collection and transfer unit is used to perform the operation of step S9 in Example 1; The multi-source information fusion unit is used to perform the operation of step S10 in embodiment 1; The attributable report generation unit is used to perform the operation of step S11 in Example 1.

[0070] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.

[0071] The present invention has the following outstanding technical effects: 1. Root Cause Analysis from a System Perspective: Addressing the systemic challenge of existing technologies that can only analyze single components in isolation and struggle to determine whether a fault is an intrinsic cause or a complication propagated throughout the system, this invention proposes a diagnostic method incorporating an equipment topology map. The equipment map provides prior knowledge of the physical topological relationships between components within the asset for the diagnostic system. Utilizing this map, this invention places the initially diagnosed fault within the context of the entire system's upstream and downstream relationships for analysis and reasoning. This enables the system to distinguish between the root cause and complications of the fault, thereby ensuring that the diagnostic conclusions can accurately trace back to the root cause of the fault.

[0072] 2. Automated Integration of Heterogeneous Information to Construct an Evidence Chain: Addressing the shortcomings of existing technologies where quantitative analysis results from machine learning models, structured relational knowledge from design drawings, and unstructured textual knowledge from technical manuals are fragmented and rely on inefficient manual integration, this invention utilizes a large language model agent as the core and designs an automated multi-source heterogeneous information fusion mechanism. This mechanism receives and parses a JSON file containing all quantitative and structured evidence, and uses this as context to initiate an enhanced knowledge retrieval from an unstructured knowledge base to obtain relevant textual evidence. Finally, the agent intelligently integrates these three types of heterogeneous information sources, automatically "weaving" them into a logically rigorous diagnostic report with a complete evidence chain, thus solving the problem of missing evidence chains caused by information silos.

[0073] 3. Dynamic Planning for Diagnostic Paths Overcomes Logical Rigidity: Addressing the shortcomings of existing technologies, such as rigid diagnostic logic and inflexible processes, which cannot handle uncertain information, this invention utilizes a "diagnostic strategist" large language model agent to achieve dynamic strategy planning capabilities. When the underlying machine learning model returns a diagnostic result with insufficient confidence, traditional fixed if-then rules will fail. However, the agent in this invention can understand this uncertainty and, using its reasoning ability, dynamically generate a corrected, temporary diagnostic execution strategy, thus exhibiting extremely high intelligent robustness.

[0074] 4. Cross-Document Joint Reasoning Solves Knowledge Silos: Addressing the shortcomings of existing technologies where unstructured knowledge sources are singular and fragmented, failing to uncover implicit connections between components, this invention proposes an enhanced retrieval method guided by a device graph. Upon receiving a preliminary diagnostic conclusion, it first queries the device graph to obtain its associated components and then leverages the semantic understanding capabilities of a large model to dynamically and in parallel perform joint queries across the technical documents of multiple associated components. This method can discover and reason about complex fault modes hidden between independent documents, describing the interactions between components, thus overcoming the analytical blind spots caused by singular knowledge sources.

Claims

1. A multi-source information fusion fault diagnosis method combining equipment maps and AI models, characterized in that, Includes an offline preparation phase and an online diagnostic phase: The offline preparation phase includes the following steps: Step 1: Prepare a historical dataset of industrial assets and construct a knowledge graph of equipment topology; Step 2: Based on the historical dataset of industrial assets, train a dual machine learning diagnostic engine that includes a "diagnostician" model and a "health assessment" model; Step 3: Based on technical documents, construct an unstructured knowledge base; The online diagnostic phase includes the following steps: Step 4: Monitor the output of the "diagnostician" model. When the state changes from normal to abnormal, trigger the diagnostic process to obtain a preliminary diagnostic result. Step 5: Based on the preliminary diagnostic results, activate the "Diagnostic Strategist" agent to query the device topology knowledge graph, thereby obtaining the set of associated components of the faulty component; Step 6: For the set of associated components, call the "health assessment" model to perform a health assessment on the associated components and obtain a quantitative health index; Step 7: The "Diagnostic Strategist" agent receives the quantitative health index, compares it with the overall health status of the assets through knowledge retrieval, and performs logical reasoning. Step 8: When there is conflict in the evidence, the "diagnostic strategist" agent dynamically generates a modified diagnostic execution strategy and performs iterative reasoning until the root cause component is identified, and finally obtains the root cause reasoning conclusion, thereby achieving clear fault location. Step 9: The "diagnostic strategist" agent packages the preliminary diagnostic results, the set of related components, and the root cause reasoning conclusions into a structured evidence archive. Step 10: Activate the "Attribution Analyzer" agent, which transforms the structured evidence archive into a deep analysis report; Step 11: The "attribution analyzer" agent uses the deep analysis report as the final context and outputs a professional diagnostic report according to the attribution rules. The professional diagnostic report includes analysis of the current root cause and operational suggestions.

2. The fault diagnosis method combining equipment maps and AI models according to claim 1, Its features are: in, In step 1, the historical dataset of industrial assets includes multivariate time-series data, health status data labels, and technical documents. The multivariate time series data X is a three-dimensional array In the formula, N represents the total number of asset units. Where D is the maximum runtime, and D is the dimension of the operating signal characteristics of each unit; The health status data label For each asset unit in the training set at each time point t, a corresponding health status label is assigned. This label represents a quantitative health index, remaining lifetime, or a qualitative failure mode classification. Technical documents consist of various unstructured or semi-structured text materials related to the asset; this data is used to construct an unstructured knowledge base. Device topology knowledge graph The construction method is as follows: First, extract the core physical components based on the publicly available design drawings, P&ID diagrams, or equipment lists of the industrial assets. V is the node set of the knowledge graph; then, based on the physical flow, spatial relationships, or logical dependencies between components, a directed edge set E is defined, where each directed edge... Representation Component It is a component The upstream of or directly affected by it, and the specific relationship is recorded in natural language.

3. The multi-source information fusion fault diagnosis method combining equipment maps and AI models according to claim 1, characterized in that: in, In step 2, the dual machine learning diagnostic engine is deployed in the backend API service, including a "diagnostic expert" model and a "health assessment" model. "Diagnostician" model This is a multi-class classification model used for preliminary fault identification and localization, which uses a certain time t or time window as a reference. Key single-device D-dimensional feature vector or all devices 3D feature vector As input, the output is the health status of a specific asset or the overall asset at that moment, and the posterior probability distribution of belonging to K+1 categories. The K+1 categories include K known abnormal states and 1 normal state. "Health Assessment" Model This refers to one or more time series prediction models that provide a quantitative and probabilistic assessment of the future health status of equipment components, using a D-dimensional feature sequence of a component over W past time steps. As input, its output It is any one or a combination of two of the continuous health indicator prediction values ​​and the probability distribution vector of future failures.

4. The fault diagnosis method combining equipment maps and AI models according to claim 1, characterized in that: in, In step 4, the process of obtaining the preliminary diagnostic results is as follows: The monitoring module will collect feature vectors in real time at a preset frequency. Transmitted to the "Diagnostician" model Perform a state check when the model's highest probability output first changes from the "normal state" category to an "abnormal state" category. When the system detects a fault, it formally triggers a full-process diagnostic process. The trigger signal is defined as a tuple. In the formula, n represents the asset ID and t represents the trigger time. This represents the preliminary diagnosis results. In step 5, the process of obtaining the associated component set of the faulty component is as follows: Once the "diagnostic strategist" is activated, it receives a trigger signal. It comes from Extract the faulty component name And use it as a parameter, combined with specific fault results Call the device map The query tool performs a graph traversal algorithm and returns the results related to the faulty node. Directly related upstream component set and downstream component sets .

5. The multi-source information fusion fault diagnosis method combining equipment maps and AI models according to claim 1, characterized in that: in, In step 6, the process of obtaining the quantitative health index is as follows: The "diagnostic strategist" agent targets the set of key related components returned in the previous step. and Call the "health assessment" model Conduct independent health assessments for each upstream component. The system extracts its relevant feature sequences and inputs them into the system. To obtain an independent quantitative health index , In step 7, the logical reasoning process is as follows: The "diagnostic strategist" big language model agent receives the quantitative health index returned in step 6. In conjunction with the external multimodal database, knowledge retrieval is used to compare the data with the overall health status of the assets and perform logical reasoning. Its core task is to comprehensively evaluate the heterogeneous evidence collected in the preceding steps. The evidence includes at least qualitative labels for preliminary diagnosis, structured relationships in the knowledge graph, and quantitative values ​​for health assessment.

6. The multi-source information fusion fault diagnosis method combining equipment maps and AI models according to claim 1, characterized in that: in, In step 8, when the system identifies logical contradictions or insufficient confidence levels between different sources of evidence, the "diagnostic strategist" agent identifies such abnormal states and uses its reasoning ability to dynamically generate a corrected diagnostic execution strategy to obtain new evidence from the system. This corrected strategy may involve calling additional diagnostic tools for further data collection or marking the conflict in the final evidence archive. The above dynamic diagnostic execution strategy is an iterative process, specifically as follows: After each round of evidence collection, the "diagnostic strategist" big language model agent evaluates the current evidence archive, which includes qualitative, quantitative, and structured evidence. The iterative reasoning process continues until the "diagnostic strategist" agent determines that the evidence meets preset conditions and then stops, confirming the root cause component.

7. The fault diagnosis method combining equipment maps and AI models according to claim 1, Its features are: in, In step 9, the "diagnostic strategist" large model agent uses the preliminary diagnostic results obtained in step 2. The list of associated components obtained in step 5 The root cause reasoning conclusions, as confirmed in step 8, are packaged into a structured JSON evidence file. , In step 10, the process of generating the in-depth analysis report is as follows: The "attribution analyzer" agent transforms the structured evidence archive into a deep analysis report, a process that strictly follows a two-stage analysis and generation workflow: Phase 1, Preliminary Evidence Extraction and Fact Statement: The "attribution analyzer" agent first receives the structured evidence file. As internal evidence, a preliminary fact report is generated by paraphrasing and summarizing the objective facts contained therein. In this process, the intelligent agent does not introduce external knowledge, make subjective inferences, or establish causal relationships not explicitly stated in the evidence. Phase Two, Enhanced Knowledge Retrieval and Comprehensive Analysis: After completing the factual statement, the "attribution analyzer" agent first dynamically generates an enhanced query based on the content of the "preliminary fact report". ; Subsequently, the large model agent uses this query to initiate a vector retrieval from the unstructured knowledge base to obtain relevant external knowledge.

8. The multi-source information fusion fault diagnosis method combining equipment maps and AI models according to claim 1, characterized in that: in, In step 11, the attribution rule is: Every statement must be clearly attributed to its source; when evidence is insufficient, it must be stated truthfully without fabricating answers; all causal relationships must be based on what is clearly stated in internal evidence or external knowledge without creating new causal chains.

9. A multi-source information fusion fault diagnosis system combining equipment maps and AI models, characterized in that, Includes offline preparation and online diagnostic modules: The offline preparation module is used to perform the following operations: Construct a device topology knowledge graph; Based on historical industrial asset datasets, a dual machine learning diagnostic engine was trained, comprising a "diagnostician" model and a "health assessment" model. An unstructured knowledge base is built based on industrial asset technical documents; The online diagnostic module is connected to the offline preparation module and includes the following components: The model-driven fault triggering unit monitors the output of the "diagnostician" model and triggers the diagnostic process when the state changes from normal to abnormal to obtain preliminary diagnostic results. The "Diagnostic Strategist" agent activation and correlation analysis unit, based on the preliminary diagnostic results, activates the "Diagnostic Strategist" agent to query the device topology knowledge graph, thereby obtaining the associated component set of the faulty component; The evidence collection unit for cross-validation calls the "health assessment" model to perform a health assessment on the associated components for the set of associated components, and obtains a quantitative health index; The cross-validated logical reasoning unit, the "diagnostic strategist" agent, receives the quantified health index, compares it with the overall health status of the assets through knowledge retrieval, and performs logical reasoning. The iterative reasoning and root cause confirmation unit, when there is conflict in the evidence, the "diagnostic strategist" agent dynamically generates a modified diagnostic execution strategy and performs iterative reasoning until the root cause component is confirmed, and finally obtains the root cause reasoning conclusion, thereby achieving clear fault location. In the evidence collection and transfer unit, the "diagnostic strategist" agent packages the preliminary diagnostic results, the set of related components, and the root cause reasoning conclusions to generate a structured evidence archive. The multi-source information fusion unit activates the "attribution analyzer" intelligent agent, which transforms the structured evidence archive into a deep analysis report. The attributable report generation unit uses the deep analysis report as the final context and outputs a professional diagnostic report according to the attribution rules. The professional diagnostic report includes analysis of the current root cause and operational suggestions.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

Images