A method and system for data processing and insights based on multi-modal intelligent assistants

Abstract

This application discloses a data processing and insight method and system based on a multimodal intelligent assistant, belonging to the field of computer technology. By integrating device physical model deduction and domain knowledge graph retrieval, a dual-track coupled reasoning framework is constructed, effectively solving the problems of decoupling between evidence and the spatiotemporal evolution of the device, and the logical defocusing of knowledge retrieval in traditional diagnosis. Through heterogeneous intelligent agent collaborative data collection and real-time conflict detection, the reliability and consistency of evidence are improved. Dynamic graph fusion and traceable reasoning mechanisms provide a transparent causal attribution chain and comprehensive credibility assessment, overcoming the limitations of black-box diagnosis. The introduction of metacognitive reflection and closed-loop optimization instructions can proactively address uncertainty, achieve continuous knowledge evolution and self-improvement of diagnostic capabilities, thereby reducing the false positive and false negative rates and supporting accurate operation and maintenance decisions.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing and insight method and system based on a multimodal intelligent assistant. Background Technology

[0002] With the popularization of Industrial Internet of Things (IIoT) and artificial intelligence (AI) technologies, intelligent equipment diagnostic systems have evolved from relying on single-source data to multimodal data fusion. However, related technical solutions still face interconnected systemic bottlenecks in achieving reliable, interpretable, and continuously evolving deep fault insights.

[0003] First, at the multimodal fusion level, most related technologies employ feature stitching or post-fusion strategies, failing to strongly bind vibration, image, and other data with the three-dimensional entity of the equipment and its spatiotemporal evolution. This results in "floating" evidence, and knowledge retrieval is prone to deviating from specific physical fault propagation paths, exhibiting "physical decoupling" and "logical defocusing" problems. Second, they rely heavily on static knowledge bases or rules, making it difficult to adapt to dynamic scenarios such as equipment aging and new types of faults. Knowledge updates are lagging, and the diagnostic process is often based on a "black box" model, lacking a transparent causal reasoning chain that aligns with engineering thinking, leading to a single, unreliable conclusion. Finally, most related technologies stop at outputting diagnostic conclusions. When evidence conflicts or conclusions are highly uncertain, they lack the ability to proactively plan and execute supplementary verification, and cannot form a closed loop of verification feedback to drive system self-optimization.

[0004] These defects combine to cause existing systems to have technical problems when dealing with complex and compound faults, such as high rates of false alarms and missed detections, unauditable reasoning processes, and conclusions that are difficult to directly support accurate closed-loop operation and maintenance decisions. Summary of the Invention

[0005] This application provides a data processing and insight method and system based on a multimodal intelligent assistant, the technical solution of which is as follows: On the one hand, a data processing and insight method based on a multimodal intelligent assistant is provided, the method comprising: The physical model of the equipment is used to deduce the fault mechanism to obtain the physical fault prior set, and the domain knowledge graph is used to perform contextual retrieval to obtain the path of the associated knowledge subgraph; The physical fault prior set and the associated knowledge subgraph path are coupled in a dual-track manner and constraint injection is performed to generate a reasoning skeleton. The reasoning skeleton defines the fault hypothesis space to be verified, the multimodal evidence requirements, and the embedded physical and spatiotemporal constraints. Based on the inference framework, multiple heterogeneous intelligent agents are scheduled to perform collaborative evidence collection, and conflict detection and cross-validation are performed based on the evidence acquired in real time during the collection process to form a structured evidence network with consistency tags; The structured evidence network and the associated knowledge subgraph path are dynamically fused and traceable reasoned to generate a comprehensive credibility assessment and an explainable attribution chain. Based on the comprehensive credibility assessment and the consistency marker, metacognitive reflection is triggered to generate a closed-loop optimization instruction. The closed-loop optimization instruction is used to perform supplementary verification or update the domain knowledge graph and the device physical model.

[0006] On the one hand, a data processing and insight system based on a multimodal intelligent assistant is provided, the system comprising: The deduction and retrieval module is used to deduce the fault mechanism of the equipment physical model to obtain the physical fault prior set, and to perform contextual retrieval of the domain knowledge graph to obtain the path of the associated knowledge subgraph; The skeleton generation module is used to perform dual-track coupling and constraint injection on the physical fault prior set and the associated knowledge subgraph path to generate a reasoning skeleton. The reasoning skeleton defines the fault hypothesis space to be verified, the multimodal evidence requirements, and the embedded physical and spatiotemporal constraints. The verification module is used to schedule multiple heterogeneous intelligent agents to perform collaborative evidence collection based on the inference skeleton, and to perform conflict detection and cross-verification based on the evidence acquired in real time during the collection process, forming a structured evidence network with consistency tags. The instruction generation module is used to perform dynamic graph fusion and traceable reasoning on the structured evidence network and the associated knowledge subgraph path to generate a comprehensive credibility assessment and an interpretable attribution chain. Based on the comprehensive credibility assessment and the consistency mark, it triggers metacognitive reflection to generate a closed-loop optimization instruction. The closed-loop optimization instruction is used to perform supplementary verification or update the domain knowledge graph and the device physical model.

[0007] On one hand, a computer device is provided, the computer device including one or more processors and one or more memories, the one or more memories storing at least one computer program, the computer program being loaded and executed by the one or more processors to implement the data processing and insight method based on a multimodal intelligent assistant.

[0008] On the one hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to implement the data processing and insight method based on a multimodal intelligent assistant.

[0009] On the one hand, a computer program product or computer program is provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the aforementioned data processing and insight method based on a multimodal intelligent assistant. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the implementation environment of a data processing and insight method based on a multimodal intelligent assistant provided in an embodiment of this application; Figure 2 This is a flowchart of a data processing and insight method based on a multimodal intelligent assistant provided in an embodiment of this application; Figure 3 This is a flowchart of another data processing and insight method based on a multimodal intelligent assistant provided in an embodiment of this application; Figure 4 This is a flowchart of another data processing and insight method based on a multimodal intelligent assistant provided in the embodiments of this application; Figure 5 This is a flowchart of another data processing and insight method based on a multimodal intelligent assistant provided in the embodiments of this application; Figure 6 This is a flowchart of another data processing and insight method based on a multimodal intelligent assistant provided in the embodiments of this application; Figure 7 This is a flowchart of another data processing and insight method based on a multimodal intelligent assistant provided in the embodiments of this application; Figure 8 This is a flowchart of another data processing and insight method based on a multimodal intelligent assistant provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of a data processing and insight system based on a multimodal intelligent assistant provided in an embodiment of this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0013] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0014] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0015] Figure 1 This is a schematic diagram illustrating the implementation environment of a data processing and insight method based on a multimodal intelligent assistant provided in an embodiment of this application. See also... Figure 1 The implementation environment may include computer equipment 110 and system 140.

[0016] Computer device 110 is connected to system 140 via a wireless or wired network. Optionally, computer device 110 may be a smartphone, tablet, laptop, desktop computer, etc., but is not limited to these. Computer device 110 has applications installed and running that support data processing and insights based on a multimodal intelligent assistant.

[0017] System 140 is a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. System 140 can provide background services for applications running on computer device 110.

[0018] Traditional intelligent equipment diagnostic systems often suffer from physical decoupling between evidence and the actual equipment and its spatiotemporal evolution during multimodal data fusion due to feature splicing or post-fusion strategies. This leads to logical defocusing in knowledge retrieval. Furthermore, these systems rely heavily on static knowledge bases, making them ill-suited for dynamic scenarios and resulting in lagging knowledge updates. The diagnostic process lacks a transparent causal reasoning chain, leading to conclusions with limited credibility. When evidence conflicts or uncertainty is high, the system lacks proactive verification and closed-loop optimization capabilities, resulting in high false positive and false negative rates, unauditable reasoning, and difficulty in supporting accurate operational and maintenance decisions.

[0019] To address this, this application proposes a data processing and insight method based on a multimodal intelligent assistant. This method generates a reasoning framework through a dual-track coupling of fault mechanism deduction from the device's physical model and contextual retrieval from the domain knowledge graph. This framework defines the fault hypothesis space to be verified, the multimodal evidence requirements, and physical and spatiotemporal constraints. Based on the reasoning framework, heterogeneous intelligent agents are scheduled to collaboratively collect evidence, performing conflict detection and cross-validation during the collection process to form a structured evidence network with consistency markers. Furthermore, the structured evidence network is fused with a dynamic path graph of the associated knowledge subgraphs for traceable reasoning, generating a comprehensive credibility assessment and an interpretable attribution chain. Based on the comprehensive credibility assessment and consistency markers, metacognitive reflection is triggered, generating closed-loop optimization instructions to supplement or update the domain knowledge graph and the device's physical model.

[0020] For ease of understanding, the following explains some key terms in this embodiment: A physical model of equipment refers to a digital representation that uses mathematical or simulation modeling to depict the internal structure, working principle, material properties, and fault propagation mechanism of equipment. This model can simulate the behavior of equipment under different operating conditions and predict the physical response when a fault occurs.

[0021] Fault mechanism deduction refers to the process of deriving potential fault modes and their theoretical physical characteristics based on abnormal phenomena of equipment, using methods such as forward or reverse simulation and causal chain analysis, starting from the physical model of the equipment.

[0022] The physical fault prior set refers to a set of potential fault hypotheses and their corresponding theoretical physical characteristics obtained through fault mechanism deduction. This set provides a preliminary fault range for subsequent evidence collection and verification.

[0023] A domain knowledge graph is a knowledge system that organizes and stores knowledge in a specific domain (such as equipment operation and maintenance, fault diagnosis) in the form of a graph structure. This graph contains nodes and edges of various types, such as entities, relationships, and events, and can represent complex domain knowledge and experience.

[0024] Contextual retrieval refers to the process of retrieving knowledge fragments or subgraph paths that are highly relevant to the current diagnostic task from a domain knowledge graph based on information such as the current equipment operating condition and alarm signals.

[0025] The associated knowledge subgraph path refers to the set of knowledge paths retrieved from the domain knowledge graph through contextual retrieval that are relevant to the current fault diagnosis task and have specific semantic associations. This set of paths provides historical experience and fault association information.

[0026] Dual-track coupling refers to the process of fusing and aligning the set of physical fault priors derived from the physical model of the equipment with the paths of the associated knowledge subgraph derived from the domain knowledge graph. The aim is to combine physical laws and domain experience to form a more comprehensive understanding of faults.

[0027] Constraint injection refers to the process of embedding physical consistency constraints and spatiotemporal acquisition constraints extracted from the device physical model and associated knowledge subgraph paths into the inference skeleton to guide the evidence collection and inference process.

[0028] The inference skeleton is a structured framework generated after dual-track coupling and constraint injection to guide subsequent fault diagnosis procedures. This skeleton clarifies the fault hypothesis space to be verified, the required types of multimodal evidence, and the relevant physical and spatiotemporal constraints.

[0029] The fault hypothesis space to be verified refers to the set of all potential fault modes defined in the reasoning skeleton that need to be verified by evidence.

[0030] Multimodal evidence requirements refer to the set of evidence types from different sensors or data sources required to effectively verify each fault hypothesis in the fault hypothesis space to be verified.

[0031] Physical constraints refer to the limiting conditions derived from the physical model of a device to ensure that fault assumptions are consistent with physical laws. Examples include energy conservation and mass conservation.

[0032] Spatiotemporal constraints refer to the restrictions imposed on the spatial location and time window of evidence collection during evidence collection, in order to ensure that the collected evidence matches the environment and time in which the fault occurred.

[0033] Heterogeneous intelligent agents refer to intelligent units with different sensing, execution, or data processing capabilities, such as vibration sensors, infrared thermal imagers, acoustic sensors, and robotic inspection units. They can work together to complete evidence collection tasks.

[0034] Collaborative evidence collection refers to the process by which multiple heterogeneous intelligent agents, guided by a reasoning framework, jointly execute multimodal evidence collection tasks according to predetermined temporal and spatial requirements.

[0035] Conflict detection refers to the process of analyzing evidence acquired in real time during the evidence collection process to identify logical contradictions or numerical incompatibilities between different pieces of evidence.

[0036] Cross-validation refers to the process of verifying or clarifying conflicting evidence by scheduling relevant heterogeneous intelligent agents to conduct supplementary retesting.

[0037] Consistency markers are identifiers used in structured evidence networks to mark the consistency or conflict states between evidence nodes.

[0038] A structured evidence network refers to a set of evidence that is organized and represented in a graph structure, showing all the standardized evidence nodes collected and their supporting, contradictory, or verifying relationships with each other.

[0039] Dynamic graph fusion refers to the process of dynamically inserting evidence nodes from a structured evidence network as temporary nodes into the paths of related knowledge subgraphs to form a unified reasoning graph.

[0040] Traceable reasoning refers to the process of performing graph-structured reasoning on a fusion reasoning graph to identify the causal logical path from key evidence nodes to candidate failure nodes, and to present that path in an interpretable form.

[0041] Comprehensive credibility assessment refers to a comprehensive evaluation that integrates physical consistency confidence, multimodal consistency confidence, and knowledge graph support to quantify the credibility of diagnostic conclusions.

[0042] An explainable attribution chain refers to a chain of causal logical reasoning paths from evidence to failure, presented in the form of natural language text, aiming to provide transparent and understandable diagnostic evidence.

[0043] Metacognitive reflection refers to the ability of a system to examine and evaluate its own diagnostic process and conclusions based on comprehensive credibility assessment and consistency markers, and to proactively plan subsequent actions (such as supplementary verification or knowledge updates).

[0044] Closed-loop optimization instructions refer to instructions generated after metacognitive reflection to guide the system to perform self-optimization, including immediate verification instructions and long-term evolution instructions.

[0045] See Figure 2 The technical solution provided in this application includes the following steps.

[0046] 201. Perform fault mechanism deduction on the physical model of the equipment to obtain the prior set of physical faults.

[0047] For example, possible failure modes of the equipment can be predefined, and the theoretical physical characteristics of each failure mode can be manually configured in the physical model of the equipment to form a set of physical failure priors.

[0048] 202. Perform contextual retrieval on the domain knowledge graph to obtain the paths of related knowledge subgraphs.

[0049] For example, based on preset fault types or equipment component names as search keywords, a direct matching query can be performed in the domain knowledge graph to obtain knowledge subgraph paths directly associated with these keywords.

[0050] 203. The physical fault prior set and the associated knowledge subgraph path are coupled in a dual-track manner and constraint injection is performed to generate the reasoning skeleton.

[0051] Specifically, fault hypotheses in the physical fault prior set can be directly merged with relevant knowledge nodes in the associated knowledge subgraph path to form a preliminary fault hypothesis space to be verified. Simultaneously, some general physical and spatiotemporal constraints can be manually set, such as the physical limits of equipment operation and the time window for evidence collection, and these constraints can be directly attached to the preliminary fault hypotheses. This generates a reasoning skeleton that clarifies the fault hypothesis space to be verified, the required multimodal evidence types, and the embedded physical and spatiotemporal constraints. For example, for each fault hypothesis, the required evidence types, such as vibration data and temperature data, can be manually specified, and corresponding physical parameter ranges and collection time periods can be set for them.

[0052] 204. Based on the reasoning framework, schedule multiple heterogeneous intelligent agents to perform collaborative evidence collection.

[0053] For example, each heterogeneous agent can be pre-configured with the type and region of evidence it is responsible for collecting, and the corresponding agents can be activated sequentially to collect data according to the order set in the reasoning skeleton.

[0054] 205. During the collection process, conflict detection and cross-validation are performed based on the evidence acquired in real time to form a structured evidence network with consistency markers.

[0055] Specifically, the collected evidence data can be preliminarily examined. For example, a simple comparison can be made between evidence data from different modalities within the same time period. If significant differences are found, the data is marked as potentially conflicting. When a conflict is detected, the conflicting evidence can be retested. For example, operators can be instructed to use a backup sensor to measure the same location again. This forms a structured evidence network with consistency markers, where each evidence node is marked as either verified or conflicting.

[0056] 206. Dynamically fuse structured evidence networks with associated knowledge subgraph paths and perform traceable reasoning to generate a comprehensive credibility assessment and an explainable attribution chain.

[0057] For example, evidence nodes in a structured evidence network can be directly overlaid onto corresponding knowledge nodes in the associated knowledge subgraph path to form a fusion reasoning graph. On this fusion reasoning graph, logical connections from key evidence to failure hypotheses can be traced, thereby generating an interpretable attribution chain. Based on the association between evidence and knowledge in the fusion reasoning graph, and the consistency markers in the structured evidence network, a comprehensive credibility assessment can be performed on each failure hypothesis. For example, the assessment result can be given by manually judging the strength of evidence support and the degree of conflict.

[0058] 207. Based on comprehensive credibility assessment and consistency marking, metacognitive reflection is triggered to generate closed-loop optimization instructions.

[0059] For example, when the evaluation results show low credibility or unresolved conflicts, the system can prompt operators to conduct supplementary verification, or experts can manually update the domain knowledge graph or equipment physical model, thereby achieving continuous optimization of the system.

[0060] This method integrates device physical model deduction with domain knowledge graph retrieval to construct a dual-track coupled reasoning framework, effectively solving the problems of decoupling between evidence and the spatiotemporal evolution of the device, and the logical defocusing of knowledge retrieval in traditional diagnostics. Through heterogeneous intelligent agent collaborative data collection and real-time conflict detection, the reliability and consistency of evidence are improved. Dynamic graph fusion and traceable reasoning mechanisms provide a transparent causal attribution chain and comprehensive credibility assessment, overcoming the limitations of black-box diagnostics. The introduction of metacognitive reflection and closed-loop optimization instructions enables the system to proactively respond to uncertainty, achieve continuous knowledge evolution and self-improvement of diagnostic capabilities, thereby reducing false positives and false negatives and supporting accurate operational and maintenance decisions.

[0061] In some embodiments described above, this application proposes to deduce a set of physical fault priors from a physical equipment model through fault mechanism inference, and to obtain related knowledge subgraph paths from a domain knowledge graph through contextual retrieval, which serve as the basic input for generating the reasoning skeleton. However, without a refined inference and retrieval mechanism, the generated initial set of fault hypotheses may lack accuracy, and the retrieved knowledge subgraphs may deviate from the actual physical fault propagation paths. This exacerbates the decoupling problem between the physical model and the knowledge graph, leading to logical defocus and affecting the reliability and interpretability of subsequent fault diagnosis.

[0062] To address this, this application further proposes to deduce a set of physical fault priors by performing fault mechanism deduction on the equipment's physical model, and to obtain the path of the associated knowledge subgraph by performing contextual retrieval on the domain knowledge graph. See [link to relevant documentation]. Figure 3 It includes the following steps: 301. Based on the current equipment operating parameters and alarm signals, generate an initial set of fault assumptions through the physical model of the equipment, and determine the theoretical physical characteristics corresponding to the initial set of fault assumptions.

[0063] 302. Using the current equipment operating parameters and the alarm signal as the retrieval context, retrieve the initial associated knowledge subgraph from the domain knowledge graph.

[0064] 303. Based on the historical experience rules and fault associations contained in the initial association knowledge subgraph, the initial fault hypothesis set is verified and filtered to generate the physical fault prior set.

[0065] 304. Based on the prior set of physical faults, construct a structured query for the knowledge graph of this domain, perform deep retrieval of subgraph paths, and obtain the associated knowledge subgraph path.

[0066] Specifically, current equipment operating parameters and alarm signals refer to real-time or near real-time measurement data of the equipment under specific operating conditions, such as sensor data for temperature, pressure, flow rate, speed, current, and voltage, as well as abnormal event indications detected by the equipment monitoring system based on preset thresholds or abnormal modes, such as over-temperature alarms, abnormal pressure alarms, and vibration over-limit alarms. These parameters and signals can be acquired in real time through industrial control systems (such as SCADA and DCS) or IoT sensor networks, or they can be fault codes or abnormal records in the event logs of the equipment's HMI (Human Machine Interface).

[0067] A physical model of equipment is a simulation model that mathematically and logically describes the internal structure, inter-component connections, material properties, and physical laws (such as thermodynamics, fluid mechanics, and mechanical dynamics) of an equipment. This model can simulate the behavioral response of the equipment under different operating conditions and fault conditions, such as digital twin models, finite element analysis models, or first-principles simulation models. For example, a physical model of equipment can be a system dynamics model built using MATLAB / Simulink, which includes detailed physical parameters and interaction relationships of components such as gearboxes, bearings, and motors.

[0068] Generating an initial set of fault hypotheses involves using the equipment's physical model, combined with current equipment operating parameters and alarm signals, to identify a list of potential fault modes that may lead to the current anomaly through methods such as forward or reverse simulation and fault propagation path analysis. This process aims to initially delineate the scope of the fault at a physical level. For example, the abnormal physical quantity indicated by the alarm signal can be used as input to perform reverse propagation analysis in the equipment's physical model, tracing back to possible root cause faulty components or fault modes. Alternatively, a series of typical fault modes can be preset, simulated in the equipment's physical model, and the deviation between the simulation results and the actual alarm signals can be compared to select the fault mode that best matches the current anomaly as the initial fault hypothesis.

[0069] Determining the theoretical physical characteristics corresponding to the initial set of fault assumptions refers to using a physical model of the equipment to simulate or calculate each fault assumption in the initial set, predicting the observable characteristics that the equipment should exhibit under different physical modes (such as vibration, heat, sound, and electricity) when the fault occurs. These characteristics are idealized representations derived from a purely physical model. For example, for a fault assumption of bearing wear, the physical model of the equipment can simulate and predict the vibration spectrum characteristics, bearing temperature rise, and possible abnormal sound characteristics generated at a specific speed.

[0070] Retrieval context refers to the contextual information used to guide knowledge graph retrieval. It is formed by semantic parsing and structuring of current equipment operating parameters and alarm signals. Its purpose is to limit the retrieval scope to the knowledge domain most relevant to the current equipment status and abnormal phenomena, thereby improving retrieval accuracy. For example, the retrieval context can be extracted from key operating modes (such as "high-speed operation" and "heavy load") in equipment operating parameters and fault types (such as "motor overheating" and "insufficient pump flow") in alarm signals, forming a query condition containing structured fields such as equipment ID, timestamp, operating mode, fault type, and abnormal physical quantity.

[0071] A domain knowledge graph is a graph-structured knowledge base built for a specific industrial domain (such as power, petrochemical, or aerospace). It contains entities such as equipment components, failure modes, maintenance activities, sensor data, expert experience, and historical cases, along with the semantic relationships between them. These relationships can be causal, compositional, attribute-based, etc. For example, a domain knowledge graph can be an ontology knowledge base built on RDF (Resource Description Framework) or OWL (Web Ontology Language), which defines concepts such as equipment type, component hierarchy, failure symptoms, diagnostic methods, and their interrelationships.

[0072] Retrieving the initial related knowledge subgraph refers to performing a query operation in the domain knowledge graph based on the retrieval context to extract a local knowledge network that is directly or indirectly related to the current equipment operating condition and alarm signals. This subgraph contains entities, events, and relationships related to the current problem, providing knowledge support for subsequent fault hypothesis verification. For example, through keyword matching or semantic similarity calculation, the nodes most relevant to entities such as equipment parts and fault phenomena in the retrieval context can be found from the knowledge graph. Then, using these nodes as centers, a graph traversal of a certain number of hops is performed to obtain their neighboring nodes and connecting edges, forming an initial related subgraph.

[0073] Historical experience rules and fault correlations refer to rules extracted from historical fault cases, expert experience, maintenance manuals, etc., that describe the probability of fault occurrence, symptom manifestation, or diagnostic logic under specific conditions. Fault correlations refer to the causal, accompanying, or compositional relationships between entities (such as components, faults, and symptoms) in a knowledge graph, such as "bearing wear leads to abnormal vibration" or "insufficient pump flow is related to impeller corrosion." For example, a historical experience rule can be expressed as the IF-THEN rule: "If the motor current is too high and the temperature is abnormal, there may be a winding short-circuit fault."

[0074] Validation and filtering refer to the process of validating and optimizing the initial set of fault hypotheses derived from the physical model of the equipment by utilizing historical experience rules and fault associations in the initial knowledge graph. This process aims to eliminate hypotheses that are inconsistent with domain knowledge and improve the credibility of the remaining hypotheses, thereby generating a more accurate set of physical fault priors. For example, the initial fault hypotheses can be matched with fault patterns in the knowledge graph; if a hypothesis lacks corresponding supporting evidence in the knowledge graph or contradicts known rules, it is discarded.

[0075] The physical fault prior set is a list of fault hypotheses deemed to have high credibility after being verified and filtered by the domain knowledge graph. It not only includes the fault modes themselves but also their corresponding theoretical physical characteristics, serving as the starting point for subsequent inference. For example, the physical fault prior set can be a list of multiple tuples, each including (fault mode ID, fault mode description, theoretical vibration characteristics, theoretical thermal characteristics, theoretical acoustic characteristics, prior confidence level).

[0076] Building structured queries for a domain-specific knowledge graph refers to generating query statements with clear syntax and semantics that can be directly executed by a domain knowledge graph query engine, based on information from the prior set of physical faults (such as fault modes and related components). These queries aim to delve deeper into the more detailed knowledge paths related to the prior fault modes. For example, a structured query could be a SPARQL query statement used to retrieve all historical maintenance records, expert diagnostic recommendations, and possible root causes of a specific fault mode.

[0077] Deep retrieval of subgraph paths refers to the multi-hop traversal of a domain knowledge graph, starting from a set of physical fault priors and following the relationships between entities, to discover longer and more complex knowledge chains or causal paths. This deep retrieval aims to reveal the propagation mechanism of faults, their potential impact, and more comprehensive contextual information. For example, breadth-first search (BFS) or depth-first search (DFS) algorithms can be used, starting from the fault node in the set of physical fault priors and expanding along predefined relationship types until a preset depth is reached or a specific target node is found.

[0078] A knowledge graph path is a set of knowledge paths obtained through deep retrieval that are highly correlated with failure modes in the prior set of physical failures and have clear semantics and topological structure. These paths can reveal the causal chain, propagation path, related symptoms, and historical solutions of failures, providing rich knowledge context for the construction of subsequent reasoning frameworks. For example, a knowledge graph path can be a series of paths represented as "(Component A) - (Failure F) - (Symptom S) - (Diagnostic Method D)", or a causal chain of "(Failure F1) - (Cause) -> (Failure F2) - (Impact) -> (Component B)".

[0079] Through the above technical solution, this application effectively solves the problems of inaccurate and irrelevant initial diagnostic inputs, improving the reliability and interpretability of fault diagnosis. Based on the current equipment operating parameters and alarm signals, an initial set of fault hypotheses is generated through equipment physical model deduction and its theoretical physical characteristics are determined. This ensures that the generation of fault hypotheses is closely linked to the real-time operating status of the equipment, avoiding the disconnect between the physical model and the actual operating conditions, and ensuring the physical rationality of the initial hypotheses. Using the current equipment operating parameters and alarm signals as the retrieval context, an initial associated knowledge subgraph is retrieved from the domain knowledge graph. This allows knowledge retrieval to focus on the current fault scenario, effectively avoiding the "logical defocusing" problem and ensuring a high degree of relevance between the acquired knowledge and the actual physical fault propagation path. On this basis, the initial set of fault hypotheses is verified and filtered using historical experience rules and fault associations in the initial associated knowledge subgraph, generating a physical fault prior set. This not only utilizes domain expert experience and historical data to verify and correct the physical model deduction results, improving the accuracy and credibility of fault hypotheses, but also effectively addresses the problem of knowledge update lag. Based on the prior set of physical faults, a structured query is constructed and a deep search of the subgraph path is performed to obtain the associated knowledge subgraph path. This makes the query of the knowledge graph more accurate and in-depth, and can reveal a more comprehensive fault causal chain and propagation mechanism. It provides high-quality and strongly associated physical and knowledge dual-track input for the subsequent construction of the reasoning skeleton, thus laying a solid foundation for the entire fault diagnosis process and ensuring the accuracy and interpretability of the diagnostic conclusions.

[0080] In some of the solutions mentioned above in this application, an initial set of fault assumptions is generated based on the current equipment operating parameters and alarm signals to preliminarily identify faults. However, in this process, due to the lack of digital twin simulation and reverse tracing mechanism, it is impossible to accurately quantify abnormal physical quantities and deviation values, and it is also difficult to locate the root cause equipment components from the physical propagation mechanism. As a result, the initial set of fault assumptions may deviate from the actual fault propagation path, affecting the accuracy and reliability of subsequent diagnosis.

[0081] To address this, this application further proposes a method for generating an initial set of fault hypotheses based on current equipment operating parameters and alarm signals through equipment physical model deduction, and for determining the theoretical physical characteristics corresponding to the initial set of fault hypotheses. The specific steps include: inputting current equipment operating parameters into the equipment physical model, performing a digital twin simulation under normal operating conditions to obtain baseline physical response data; comparing the abnormal data indicated by the alarm signal with the baseline physical response data to identify abnormal physical quantities with significant deviations and their corresponding deviation values; tracing the abnormal physical quantities back to one or more potential root cause equipment components based on the fault propagation mechanism network built into the equipment physical model to form an initial set of fault hypotheses; and for each fault hypothesis in the initial set of fault hypotheses, injecting the corresponding fault conditions into the equipment physical model and performing simulations to predict and output the theoretical physical characteristics of each fault hypothesis under vibration, thermal imaging, and acoustic modes.

[0082] Specifically, the current equipment operating parameters are input into the equipment physical model, and a digital twin simulation under normal operating conditions is performed to obtain baseline physical response data. The current equipment operating parameters refer to various observable or measurable physical quantities of the equipment in its current operating state, such as sensor data for temperature, pressure, speed, current, and voltage, as well as configuration information such as the equipment's operating mode and load. These parameters comprehensively reflect the real-time operating status of the equipment. The equipment physical model is a precise mapping of the equipment in digital space. It is constructed through mathematical equations, physical laws, and engineering experience, and can simulate the behavior and response of the equipment under different operating conditions. The digital twin simulation uses this physical model to reproduce the equipment's operation under normal conditions in a virtual environment, thereby generating a series of expected, fault-free physical response data, i.e., baseline physical response data. This data provides an objective and quantitative reference benchmark for subsequent anomaly detection, avoiding the subjectivity and inaccuracy that may arise from relying solely on experience.

[0083] By comparing the abnormal data indicated by alarm signals with the baseline physical response data, significant deviations in physical quantities and their corresponding deviation values ​​can be identified. The abnormal data indicated by alarm signals typically originates from the device's built-in monitoring system or sensors, triggered when certain parameters exceed preset thresholds. By precisely comparing these real-time abnormal data with previously obtained baseline physical response data, the abnormal physical quantities and their degree of abnormality (i.e., deviation value) can be quantitatively identified. This comparison can be achieved through statistical methods (such as mean and variance analysis), machine learning algorithms (such as anomaly detection models), or simple threshold comparisons, thereby ensuring the accuracy and objectivity of anomaly identification.

[0084] Building upon this foundation, and based on the fault propagation mechanism network embedded in the equipment physical model, abnormal physical quantities are traced back to one or more potential root cause equipment components, forming an initial set of fault hypotheses. The fault propagation mechanism network is a crucial component of the equipment physical model; it describes the physical connections between components within the equipment, energy transfer paths, and the causal relationships of how faults propagate from one component to another. For example, it can be a fault tree, a causal graph, or a state transition diagram. By performing reverse reasoning on the identified abnormal physical quantities—that is, tracing back along the fault propagation mechanism network—the initial source causing these anomalies can be located, i.e., the root cause equipment component. For example, if an abnormally high bearing temperature is detected, the fault propagation mechanism network might trace it back to a lubrication system failure or bearing wear. The resulting initial set of fault hypotheses is based on physical mechanisms and possesses high reliability.

[0085] For each fault hypothesis in the initial set of fault hypotheses, corresponding fault conditions are injected into the equipment physical model, and simulations are performed to predict and output the theoretical physical characteristics of each fault hypothesis under vibration, thermal imaging, and acoustic modes. For each potential fault in the initial set of fault hypotheses, such as "bearing wear," the occurrence of the fault is simulated in the equipment physical model, for example, by adjusting the bearing's friction coefficient or clearance parameters. Then, digital twin simulations are performed again to predict the theoretical characteristics of the fault under different physical modes (such as the spectral characteristics of vibration signals, the temperature distribution of thermal images, and the frequency characteristics of acoustic signals). These theoretical physical characteristics provide clear targets and basis for subsequent evidence collection and verification, supporting the fusion diagnosis of multimodal data.

[0086] Through the above technical solutions, this application can accurately generate an initial set of fault hypotheses and predict their theoretical physical characteristics. By acquiring benchmark physical response data through digital twin simulation, objective and quantitative references are provided for anomaly detection, avoiding errors caused by subjective experience. Comparing the anomaly data indicated by alarm signals with benchmark data can accurately identify abnormal physical quantities and their deviations, ensuring the accuracy of anomaly identification. Furthermore, by using the fault propagation mechanism network built into the equipment physical model for reverse tracing, the generation of fault hypotheses is closely bound to the physical structure and mechanism of the equipment, effectively solving the problem of initial hypotheses deviating from the actual fault propagation path and improving the physical consistency of fault diagnosis. By injecting fault conditions into the equipment physical model and performing multimodal simulation, theoretical physical characteristics under vibration, thermal imaging, and acoustic modes are predicted for each fault hypothesis. This provides a comprehensive theoretical basis for subsequent collaborative evidence collection, enhancing the effectiveness and reliability of multimodal evidence fusion diagnosis, thus laying a solid foundation for subsequent reasoning framework construction and evidence collection.

[0087] In some of the solutions mentioned above in this application, the path of the associated knowledge subgraph is retrieved from the domain knowledge graph to support fault diagnosis. However, in this process, the retrieval may lack a deep semantic understanding of the current equipment operating parameters and alarm signals, causing the retrieval results to deviate from the specific physical fault propagation path and resulting in a logical defocusing problem.

[0088] To address this, this application proposes a method that uses current equipment operating parameters and alarm signals as the retrieval context to retrieve an initial associated knowledge subgraph from a domain knowledge graph. This method performs joint semantic understanding on the current equipment operating parameters and alarm signals to extract equipment component identifiers, fault phenomenon descriptions, and operating context information. Specifically, joint semantic understanding aims to transform raw equipment operating data and alarm information into structured, semantically rich knowledge elements. For example, Natural Language Processing (NLP) technology can be used to parse the text description of alarm signals, and combined with a predefined equipment ontology or professional dictionary, key information such as equipment component names, fault type keywords, and abnormal values ​​can be identified. Simultaneously, equipment operating parameters (such as temperature, pressure, and speed) are compared with preset thresholds or normal ranges, and parameters exceeding the range, along with their corresponding components, time, and other information, are extracted in a structured manner. Another approach is to use deep learning models, such as a Transformer-based semantic understanding model, to jointly encode equipment operating parameters (after feature engineering) and alarm signal text, generating high-dimensional semantic vectors. These semantic vectors are then mapped to the entity and relation space of the knowledge graph through a pre-trained knowledge graph embedding model, thereby automatically identifying and extracting the equipment component identifiers, fault phenomenon descriptions, and operating context information most relevant to the current operating condition and alarm signal.

[0089] The extracted equipment component identifiers, fault descriptions, and operating context information are combined to generate graph database query statements. The purpose of this step is to transform the semantically understood information into a query language that the graph database can recognize and execute. For example, a series of graph database query templates can be predefined, such as using Cypher or SPARQL query languages. The extracted equipment component identifiers, fault descriptions, and operating context information are then populated into these templates as variables to generate specific query statements. Alternatively, queries can be constructed through a programming interface (API). The system provides a query builder that takes structured equipment component identifiers, fault descriptions, and operating context information as input and dynamically concatenates and generates query statements that conform to the graph database syntax specifications, such as complex queries that include multi-hop paths, attribute filtering, and relation type restrictions, based on internal logic and the graph database schema definition.

[0090] The generated graph database query statement is submitted to the query engine of the domain knowledge graph to perform multi-hop related subgraph retrieval based on pattern matching. This step aims to retrieve knowledge subgraphs closely related to the current fault situation from a large-scale knowledge graph through an efficient graph query mechanism. Specifically, the native query engine provided by the graph database used by the domain knowledge graph (such as Neo4j, JanusGraph, etc.) can be directly utilized. This engine can efficiently parse and execute graph query language, traverse the graph structure, and perform pattern matching and path search based on the node type, relation type, attribute conditions, and hop count limits defined in the query statement, thereby returning multi-hop related subgraphs that meet the conditions. For large-scale domain knowledge graphs, a distributed graph processing framework can also be used. The query engine converts the graph database query statement into a distributed graph algorithm task, and performs graph traversal and pattern matching operations in parallel on the cluster to process ultra-large-scale knowledge graphs containing billions of nodes and edges, and efficiently discovers multi-hop related subgraphs through optimized algorithms.

[0091] The system retrieves and integrates the entity nodes, event nodes, and connections found in the multi-hop relational subgraph from the domain knowledge graph to form an initial relational knowledge subgraph. This step structures the query results into a usable knowledge subgraph. For example, the query engine typically returns a collection of nodes and edges, which the system can directly map to in-memory graph data structures (such as adjacency lists or custom graph object models). For each returned entity node (such as device parts, fault types, sensors) and event node (such as maintenance events, alarm events), as well as their connections (such as "contains", "causes", "measures"), instantiation is performed and added to the initial relational knowledge subgraph. Alternatively, the query results are serialized to a standard format (such as JSON-LD or GraphML) and then loaded into the application's knowledge representation layer through a deserialization process. During deserialization, the entity nodes, event nodes, and their connections are parsed and validated to ensure they conform to the predefined ontology structure and data model, thus constructing a structurally complete and semantically consistent initial relational knowledge subgraph.

[0092] Through the above technical solution, this application can ensure that the initial associated knowledge subgraph retrieved from the domain knowledge graph is highly relevant to the current equipment operating condition and alarm signals, solving the problems of "physical decoupling" and "logical defocusing" in traditional retrieval methods. This precise context-aware retrieval lays a solid foundation for subsequent fault hypothesis verification, screening, and the construction of the reasoning framework, improving the accuracy and reliability of fault diagnosis.

[0093] In some of the solutions described above in this application, an initial set of fault assumptions is generated based on the physical model of the equipment. However, in this process, the initial set of fault assumptions may contain invalid or low-confidence fault assumptions, and there is a lack of effective use of historical experience rules and fault correlation relationships, resulting in an inaccurate and unreliable physical fault prior set.

[0094] To address this, this application further proposes a method for validating and filtering an initial set of fault hypotheses and generating a physical fault prior set based on historical experience rules and fault associations contained in an initial associated knowledge subgraph. The specific steps include: parsing the confidence levels of historical experience rules and the topological strength of fault associations related to the initial set of fault hypotheses from the initial associated knowledge subgraph; calculating the logical fit with historical experience rules for each fault hypothesis in the initial set of fault hypotheses, and generating a knowledge support level for each fault hypothesis by combining the confidence levels of historical experience rules; determining the contextual relevance of each fault hypothesis in the initial associated knowledge subgraph based on the topological strength of fault associations; integrating the knowledge support level and contextual relevance to determine a comprehensive evidence score for each fault hypothesis; sorting and filtering the initial set of fault hypotheses based on the comprehensive evidence score, and integrating the filtered fault hypotheses and their corresponding theoretical physical characteristics into a physical fault prior set.

[0095] The process involves retrieving the confidence levels of historical experience rules and the topological strength of fault associations from the initial knowledge graph. This aims to extract highly relevant metadata for the current fault diagnosis task from the constructed initial knowledge graph, specifically the reliability (confidence) of historical experience rules and the tightness (topological strength) of associations between faults. This information forms the basis for subsequent effective verification and filtering of the initial fault hypotheses. Specifically, a predefined graph query language (such as Cypher or SPARQL) can be used to perform pattern matching in the knowledge graph, identifying nodes or edges representing historical experience rules and retrieving their stored confidence values. Simultaneously, the topological strength of fault associations is calculated by analyzing the path length, edge type, and weight between fault nodes; for example, shorter paths and higher weights result in greater topological strength. Alternatively, a machine learning-based approach can be used to train historical data in the knowledge graph, learning a predictive model for rule confidence and a model for evaluating association strength. When a new initial knowledge graph is generated, these models are used to automatically parse and assign the corresponding confidence and topological strength values.

[0096] For each fault hypothesis in the initial set of fault hypotheses, the logical fit with historical experience rules is calculated. Combined with the confidence level of these historical experience rules, a knowledge support score is generated for each fault hypothesis. The purpose of this step is to quantify the degree of conformity between each initial fault hypothesis and known historical experience rules in the domain, taking into account the reliability of these rules themselves, thereby obtaining the strength of support for the hypothesis at the knowledge level. Logical fit measures the degree of matching between the hypothesis and the rules, while knowledge support combines this matching degree with the confidence level of the rules. Specifically, natural language processing (NLP) techniques can be used to semantically parse the fault hypotheses and historical experience rules, extracting key entities and relationships, and then matching them using a rule engine or logical reasoning system. For example, if the fault hypothesis "bearing wear" is highly semantically consistent with the historical rule "bearing wear causes abnormal vibration," then the logical fit is high. Knowledge support can be defined as the product or weighted sum of the logical fit and the confidence level of the historical experience rules. Another approach is to construct an ontology-based matching system that maps failure hypotheses and historical experience rules to a unified ontological concept. Semantic similarity or logical implication relationships between these hypotheses are then calculated using ontological reasoning as the logical fit. Combined with rule confidence, the knowledge support for each failure hypothesis is calculated using methods such as fuzzy logic or Bayesian networks.

[0097] For each fault hypothesis, based on the topological strength of fault associations, the contextual relevance of each fault hypothesis in the initial associated knowledge subgraph is determined. This step aims to evaluate the "position" and "importance" of each fault hypothesis in the initial associated knowledge subgraph, i.e., how closely it is connected to other related faults, components, events, etc., within the graph structure. Topological strength reflects the reliability or strength of fault associations, while contextual relevance uses these strengths to measure the relevance of the fault hypothesis within the entire knowledge network. Specifically, centrality measures in graph theory (such as degree centrality, betweenness centrality, proximity centrality, or eigenvector centrality) can be used to evaluate the importance of fault hypothesis nodes in the subgraph. Simultaneously, combining the topological strength of fault associations as edge weights, the weighted shortest path or reachability from the fault hypothesis node to other related nodes (such as alarm signals, operating condition parameter nodes) is calculated, and this is used as the contextual relevance. Furthermore, graph embedding techniques can be employed to map the nodes and edges in the initial associated knowledge subgraph to a low-dimensional vector space. Then, the contextual relevance is determined by calculating the similarity (such as cosine similarity) between the fault hypothesis node vector and other relevant node vectors, and combining this with the topological strength of the fault association.

[0098] The goal of this step is to integrate knowledge support and contextual relevance to determine the comprehensive evidence score for each failure hypothesis. This integration combines evaluation results obtained from both historical experience rules and knowledge graph topology to form a more comprehensive and reliable overall evaluation score for each failure hypothesis. This fusion avoids the limitations of single-dimensional evaluation and provides a more robust judgment. Specifically, a weighted average method can be used, linearly combining knowledge support and contextual relevance according to preset weights to obtain the comprehensive evidence score. For example, the comprehensive evidence score = w1 × knowledge support + w2 × contextual relevance, where w1 and w2 are adjustable parameters, and w1 + w2 = 1. Alternatively, multi-criteria decision analysis methods (such as Analytic Hierarchy Processing (AHP) and Topology-Based Solution) or machine learning models (such as Support Vector Machines (SVM) and decision trees) can be used, with knowledge support and contextual relevance as input features, to train the model to predict the comprehensive evidence score for each failure hypothesis.

[0099] The initial set of fault hypotheses is sorted and filtered based on a comprehensive evidence score. The selected fault hypotheses and their corresponding theoretical physical characteristics are then integrated into a physical fault prior set. This step is crucial for generating the physical fault prior set. By quantifying and evaluating the initial fault hypotheses, sorting and filtering ensures that only those with high confidence and relevance are included, thereby improving the efficiency and accuracy of subsequent reasoning. Specifically, a preset comprehensive evidence score threshold can be set. All fault hypotheses with scores above this threshold are retained, while those below are discarded. Alternatively, a fixed number N can be set, retaining only the N fault hypotheses with the highest scores. After filtering, these retained fault hypotheses and their theoretical physical characteristics derived from the equipment physical model are structurally integrated to form the physical fault prior set. Furthermore, statistical methods can be used, such as fitting the comprehensive evidence score to a normal distribution and then determining the filtering boundary based on confidence intervals or outlier detection methods. Alternatively, clustering algorithms can be used to classify fault hypotheses into high-confidence, medium-confidence, and low-confidence categories, and then selecting fault hypotheses from the high-confidence category.

[0100] Through the above technical solution, this application, based on the initial fault hypothesis set generated by equipment physical model deduction, further introduces historical experience rules and fault association relationships from the domain knowledge graph for dual verification and screening. Specifically, by analyzing the confidence level of historical experience rules and the topological strength of fault association relationships in the initial association knowledge subgraph, a reliable basis is provided for subsequent quantitative evaluation. For each initial fault hypothesis, its logical consistency with historical experience rules is calculated, and knowledge support is generated by combining the rule confidence level. At the same time, its contextual relevance is determined based on the topological strength of fault association relationships, thereby comprehensively evaluating the reliability of fault hypotheses from both knowledge and structural dimensions. By fusing knowledge support and contextual relevance, a comprehensive evidence score is obtained, and the initial fault hypothesis set is sorted and screened based on this score, effectively eliminating invalid or low-confidence fault hypotheses and improving the accuracy and reliability of the physical fault prior set. This not only solves the problem that the initial hypothesis may contain invalid or low-credibility content, but also provides high-quality and high-credibility prior information for subsequent reasoning skeleton generation and collaborative evidence collection, avoiding diagnostic bias and resource waste caused by inaccurate initial hypotheses, making the entire data processing and insight method more robust and efficient.

[0101] In some of the solutions mentioned above in this application, a structured query is constructed based on the physical fault prior set and a deep search is performed to obtain the path of the associated knowledge subgraph. However, in this process, the search may be inefficient, the search results may be irrelevant or incomplete, leading to deviation from the physical fault propagation path and resulting in a logical defocusing problem.

[0102] To address this, this application further proposes a method for constructing a structured query for a domain-specific knowledge graph based on the physical fault prior set, performing deep retrieval of subgraph paths, and obtaining the associated knowledge subgraph path. This method includes: determining the retrieval priority of a fault hypothesis based on a comprehensive evidence score of the fault hypothesis in the physical fault prior set; generating a parameterized graph query template based on the retrieval priority and fault mode identifiers and associated equipment component identifiers in the physical fault prior set; executing the parameterized graph query template in the domain knowledge graph, wherein retrieval paths associated with fault hypotheses with high retrieval priority are preferentially expanded to obtain a candidate subgraph path set; determining the semantic relevance between each subgraph path in the candidate subgraph path set and its corresponding fault hypothesis, and weighting and ranking them according to the semantic relevance and the retrieval priority to obtain a weighted ranking result; and selecting the top N subgraph paths based on the weighted ranking result for integration to generate the associated knowledge subgraph path, where N is a positive integer.

[0103] Specifically, determining the retrieval priority of a fault hypothesis involves ranking each fault hypothesis in the physical fault prior set based on its comprehensive evidence score. This clarifies which fault hypotheses should be prioritized during subsequent knowledge graph retrieval. For example, a threshold can be set to mark fault hypotheses with comprehensive evidence scores above this threshold as high priority, or a full ranking can be performed directly based on scores from highest to lowest. Another approach is to divide fault hypotheses into several priority levels, such as "high," "medium," and "low," based on their comprehensive evidence scores to facilitate hierarchical retrieval.

[0104] Generating parameterized graph query templates refers to constructing a dynamically populated graph query statement structure based on predetermined retrieval priorities, fault mode identifiers, and associated device component identifiers. This template guides the knowledge graph retrieval engine, enabling it to focus on knowledge paths related to specific fault modes and device components during query execution. For example, a Cypher query template can be constructed, containing placeholders for matching fault mode nodes, device component nodes, and specific relationship types between them, dynamically adjusting query depth or breadth parameters based on retrieval priorities. Another implementation approach is to utilize the SPARQL query language to construct a query template with variables that can be populated with specific fault mode URIs and device component URIs, while embedding priority-related filtering conditions or sorting instructions within the query template.

[0105] Executing this parameterized graph query template within the domain knowledge graph involves prioritizing the expansion of retrieval paths associated with high-priority failure hypotheses to obtain a set of candidate subgraph paths. This means submitting the generated parameterized graph query template to the domain knowledge graph's query engine and intelligently adjusting the graph traversal or search strategy based on the failure hypothesis's retrieval priority. For high-priority failure hypotheses, the query engine allocates more computational resources, such as increasing search depth, broadening search breadth, or employing more complex path discovery algorithms, to ensure comprehensive and efficient discovery of knowledge paths related to these high-priority failure hypotheses. For example, breadth-first search (BFS) or depth-first search (DFS) algorithms can be used, with larger search depth limits or longer search times set for high-priority failure hypotheses. Another implementation approach is to leverage the graph database's indexing optimization capabilities to pre-index nodes and edges related to high-priority failure hypotheses, thereby accelerating the expansion of their associated paths and dynamically adjusting the query engine's parallelism to prioritize high-priority queries.

[0106] The process involves determining the semantic relevance between each subgraph path in the candidate subgraph path set and its corresponding fault hypothesis, and then weighting and ranking them based on this semantic relevance and retrieval priority. This weighted ranking result means evaluating the semantic matching degree between the content of each candidate subgraph path obtained through deep retrieval and its associated fault hypothesis, and combining this with the retrieval priority of the fault hypothesis itself for a comprehensive ranking. For example, semantic relevance can be obtained by calculating the word embedding similarity, topic model similarity, or ontology concept matching degree between entities, relations, and the fault hypothesis description text in the subgraph path. During weighted ranking, linear weighted sum, product weighting, or machine learning model-based methods can be used to fuse semantic relevance with retrieval priority to generate a comprehensive score. Another implementation method is to use a graph neural network (GNN) to extract features from candidate subgraph paths and train a classifier or regressor to predict their semantic relevance with the fault hypothesis. Then, the predicted relevance and retrieval priority are weighted and summed using preset weighting factors to obtain the weighted ranking result.

[0107] Based on the weighted ranking result, the top N subgraph paths are selected and integrated to generate the associated knowledge subgraph path. N is a positive integer, representing the N highest-scoring paths selected from the candidate subgraph path set based on the weighted ranking result, and then merged or connected to form the associated knowledge subgraph path. The value of N can be dynamically adjusted according to the needs of the actual application scenario, computing resource limitations, or user configuration. For example, N can be set to a fixed value, such as N=5 or N=10, to ensure sufficient but not overly redundant knowledge paths are obtained. Another implementation method is to dynamically determine the value of N based on the distribution of the weighted ranking result. For example, all paths with scores higher than a certain dynamic threshold can be selected, or paths with cumulative scores reaching a certain proportion can be selected to ensure that the selected paths have both high relevance and a certain degree of coverage.

[0108] Through the above technical solutions, this application effectively addresses the problems of inefficiency, irrelevant or incomplete results, and logical defocusing that may occur during knowledge retrieval. By determining retrieval priorities based on a comprehensive evidence score of fault hypotheses, the system can prioritize high-confidence fault hypotheses that have been verified by the physical model and the initial knowledge graph, thereby avoiding ineffective deep searches of low-confidence hypotheses, improving retrieval efficiency and saving computational resources. Furthermore, by combining retrieval priorities, fault mode identifiers, and associated device component identifiers to generate parameterized graph query templates, the deep retrieval of the knowledge graph becomes more targeted, ensuring that the retrieval results are closely related to specific physical fault paths, effectively avoiding the problem of "logical defocusing." When executing graph queries, prioritizing the expansion of paths associated with high-priority fault hypotheses enables the efficient acquisition of a high-quality set of candidate subgraph paths. By evaluating the semantic relevance between candidate subgraph paths and fault hypotheses and weighting them according to retrieval priority, it is ensured that the selected related knowledge subgraph paths are not only highly relevant semantically, but also given priority in terms of physical credibility, thus providing more accurate and reliable knowledge support for the subsequent construction of the reasoning framework.

[0109] In some of the solutions mentioned above in this application, dual-track coupling and constraint injection are proposed to generate the reasoning skeleton. However, in its implementation, the coupling between the physical fault prior set and the associated knowledge subgraph path may not be tight enough, resulting in a lack of precise mapping between fault modes and knowledge nodes. Simultaneously, unclear or insufficient constraint injection may lead to insufficient integration of theoretical physical characteristics with historical case context, and inconsistencies in the fault hypothesis space to be verified. Furthermore, the lack of precise assignment of multimodal evidence requirements and the unstructured embedding of physical consistency constraints and spatiotemporal acquisition constraints render the reasoning skeleton lacking in executability and reliability, affecting the credibility of subsequent evidence collection and diagnosis.

[0110] To address this, this application proposes a method for generating a reasoning skeleton by dual-track coupling and constraint injection of the physical fault prior set and the associated knowledge subgraph path. (See [link to relevant documentation]). Figure 4 The method includes the following steps: 401. Perform semantic alignment and association matching between the fault modes in the physical fault prior set and the fault nodes in the associated knowledge subgraph path to construct a mapping relationship network between fault modes and knowledge nodes.

[0111] 402. Based on the mapping relationship network, the theoretical physical features in the physical fault prior set and the historical case context in the related knowledge subgraph path are fused and complementary to generate a unified fault hypothesis space to be verified, and the required multimodal evidence type is assigned to each fault hypothesis in the fault hypothesis space to be verified.

[0112] 403. Derive the physical consistency constraints corresponding to the fault hypothesis space to be verified from the deduction logic of the equipment physical model, and derive the spatiotemporal acquisition constraints from the equipment topology of the associated knowledge subgraph path.

[0113] 404. Physical consistency constraints and spatiotemporal acquisition constraints are used as structured metadata and associated with the corresponding fault hypotheses and multimodal evidence types in the fault hypothesis space to be verified, respectively, and assembled to generate a reasoning skeleton.

[0114] Specifically, when semantically aligning and matching fault patterns in the physical fault prior set with fault nodes in the associated knowledge subgraph path to construct a mapping network between fault patterns and knowledge nodes, the aim is to establish a precise correspondence between fault hypotheses derived from the physical model and empirical knowledge in the domain knowledge graph. This can be achieved in several ways. For example, Natural Language Processing (NLP) techniques can be used to calculate the semantic similarity between the textual description of the fault pattern and the labels and attributes of the fault nodes in the knowledge graph, identifying synonymous, near-synonymous, or causally related entities to achieve semantic alignment. Another approach is to employ ontology-based matching algorithms to map physical fault patterns to predefined fault ontology concepts in the domain knowledge graph, establishing a mapping network through hierarchical relationships and attribute associations between ontology.

[0115] In this study, based on a mapping network, the theoretical physical features of a physical fault prior set are fused and complementaryly corrected with historical case contexts in the associated knowledge subgraph paths to generate a unified space of fault hypotheses to be verified. Then, the required multimodal evidence types are assigned to each fault hypothesis in this space. The fusion and complementary correction aim to integrate theoretical predictions with historical experience data to form a more comprehensive and accurate description of the fault hypotheses. This can be achieved, for example, through statistical methods, comparing the parameter distributions of theoretical physical features with the distributions of actual observational data in historical case contexts to calibrate the theoretical range, or using historical data to supplement feature dimensions not considered in the theoretical model. Another approach is to use a rule-based expert system. Based on the correspondences identified by the mapping network, domain expert knowledge or machine learning models can be used to judge and correct inconsistencies between theoretical features and historical cases. For example, when there is a significant deviation between theoretical predictions and historical experience, verified historical experience data is prioritized for correction. Assigning the required multimodal evidence types to each fault hypothesis in the space of fault hypotheses to be verified aims to clarify the objectives of subsequent evidence collection. This can be achieved through a pre-defined rule base that associates specific fault modes or their physical manifestations with required sensor modalities (such as vibration, thermal imaging, acoustics, current, etc.). For example, for bearing wear faults, vibration and acoustic data might be needed. For electrical overload faults, current and thermal imaging data might be needed. Alternatively, machine learning models can be used to learn and predict the most important combination of evidence types needed for the current fault hypothesis by analyzing the valid evidence modalities relied upon by different fault types in historical diagnostic cases.

[0116] When deriving physical consistency constraints corresponding to the fault hypothesis space to be verified from the derivation logic of the equipment physical model, and deriving spatiotemporal acquisition constraints from the equipment topology of the associated knowledge subgraph path, the purpose of deriving physical consistency constraints is to ensure that the evidence subsequently collected conforms to basic physical laws. This can be achieved by parsing the physical equations, conservation laws, or state transition rules built into the equipment physical model. For example, for a rotating machine, there is a fixed physical relationship between its vibration frequency and rotational speed, which is a physical consistency constraint. Another approach is to simulate different fault conditions in the equipment physical model, observe and record the interdependencies and numerical ranges between various physical quantities, and use these relationships as physical consistency constraints. The purpose of deriving spatiotemporal acquisition constraints is to guide the actual operation of evidence collection, ensuring the timeliness and spatial relevance of the collection. This can be achieved by analyzing the spatial positional relationships of equipment components in the associated knowledge subgraph path, the fault propagation path, and the time window for collecting effective evidence in historical fault cases. For example, if the knowledge graph shows that a fault usually occurs in a specific component and its impact propagates along a specific path, then evidence collection should prioritize these components and their propagation paths, and be carried out within a specific time window after the fault occurs. Another approach is to determine the minimum spatial area and maximum collection time required for each fault hypothesis based on the device topology, combined with the sensor deployment location and coverage area.

[0117] When physical consistency constraints and spatiotemporal acquisition constraints are used as structured metadata and associated with corresponding fault hypotheses and multimodal evidence types in the fault hypothesis space to be verified, the aim is to integrate all key information into an executable diagnostic plan. This can be achieved by encapsulating these constraints and relationships using standardized data formats such as JSON or XML. For example, each fault hypothesis can be treated as an independent structured object, containing its unique identifier, theoretical physical characteristics, a list of assigned multimodal evidence types, and the physical consistency constraints directly associated with that fault hypothesis and the spatiotemporal acquisition constraints associated with each evidence type. Alternatively, graph database technology can be used to treat fault hypotheses, evidence types, physical consistency constraints, and spatiotemporal acquisition constraints as nodes in a graph, representing the relationships between them by defining different types of edges, thus forming a queryable and traversable inference skeleton graph structure.

[0118] Through the above technical solution, this application effectively solves the problems of insufficient coupling and unclear constraints between the physical fault prior set and the associated knowledge subgraph path. By semantically aligning and matching the fault modes in the physical fault prior set with the fault nodes in the associated knowledge subgraph path, a mapping network between fault modes and knowledge nodes is constructed, thereby establishing a precise connection between physical and knowledge tracks and avoiding loose coupling caused by semantic deviations. Based on this mapping network, the theoretical physical characteristics and historical case context are fused and complementaryly corrected to generate a unified fault hypothesis space to be verified, and the required multimodal evidence type is precisely assigned to each fault hypothesis. This not only solves the problem of insufficient fusion of theoretical and empirical data, but also makes the subsequent evidence collection target clearer and avoids logical defocus. Furthermore, physical consistency constraints are derived from the equipment physical model deduction logic, and spatiotemporal collection constraints are derived from the equipment topology of the associated knowledge subgraph path, ensuring the scientificity and feasibility of the constraints. These constraints are used as structured metadata, associated with the corresponding fault hypotheses and multimodal evidence types, and assembled to generate a reasoning skeleton, which greatly improves the executability and reliability of the reasoning skeleton. This reasoning framework serves as a guide for subsequent collaborative evidence collection, ensuring that the collected evidence is more targeted and effective. It also provides a solid foundation for subsequent conflict detection, cross-validation, dynamic graph fusion, and traceable reasoning, thereby improving the credibility and diagnostic efficiency of the entire data processing and insight method.

[0119] In some of the embodiments described above in this application, a unified hypothesis space for unverified faults is proposed to be generated by fusing theoretical physical features and historical case contexts based on a mapping relationship network. However, in the process of its implementation, there may be statistical deviations between theoretical features and actual observations, feature dimensions may be missing, and descriptions from different orbits may conflict, resulting in inaccurate or inconsistent hypothesis spaces, which affects the credibility and interpretability of subsequent diagnoses.

[0120] To address this, this application further proposes a method for generating a reasoning skeleton by dual-track coupling and constraint injection of the physical fault prior set and the associated knowledge subgraph path. Specifically, based on the mapping network, the theoretical physical features in the physical fault prior set and the historical case context in the associated knowledge subgraph path are fused and complementaryly corrected to generate a unified fault hypothesis space to be verified. This fusion and complementary correction process specifically includes the following steps: For each fault mode associated in the mapping network, the system obtains the corresponding theoretical physical characteristics and the set of actual observed features from the context of that historical case. In this step, the system queries the pre-established mapping network to extract the theoretical physical characteristics of each identified fault mode from the equipment physical model simulation database, such as vibration frequency, temperature curve, and pressure fluctuation range. Simultaneously, the system parses the set of actual observed features from the historical case nodes associated with the fault mode in the domain knowledge graph, such as sensor readings in historical fault reports, expert diagnostic records, and fault phenomenon descriptions in maintenance logs. Alternatively, the platform uses the fault mode identifier as a key to retrieve the corresponding theoretical physical feature vector in the physical model simulation result storage module. Then, in the knowledge graph query interface, it aggregates the actual observed feature data of all relevant historical cases by traversing the historical case edges connected to the fault mode nodes, forming a set.

[0121] A statistical difference analysis is performed between the theoretical physical characteristic representation and the actual observed characteristic set to obtain the analysis results. Based on these results, the parameter range of the theoretical physical characteristic representation is calibrated, and missing feature dimensions are supplemented from the actual observed characteristic set to obtain the calibrated and supplemented feature descriptions. Specifically, the system can use statistical methods, such as mean, variance, and distribution differences (e.g., Kolmogorov-Smirnov test), to compare the parameter distribution of the theoretical physical characteristic representation with the distribution of the actual observed characteristic set to identify parameter deviations. Simultaneously, by comparing the dimensions of feature vectors, feature dimensions present in actual observations but not directly output by the theoretical model are identified. Another approach is to use anomaly detection algorithms from machine learning, using the theoretical physical characteristic representation as a benchmark to detect deviations in the actual observed characteristic set and quantify the degree of parameter deviation. Missing feature dimensions are identified through set difference operations or sparsity analysis of the feature matrix. Based on these analysis results, the system adjusts the parameter range of theoretical physical characteristics using linear or nonlinear mapping functions according to the identified parameter deviations. For example, if the theoretical temperature range is [80, 90]℃, while actual observations are concentrated in [85, 95]℃, the theoretical range is shifted upwards by 5℃. Alternatively, methods such as Bayesian updates or Kalman filtering are used, with actual observation data as evidence, to iteratively correct the parameter range of theoretical physical characteristics. For identified missing feature dimensions, the system can extract historical data for that dimension from the actual observation feature set and calculate its statistical characteristics (such as average value, typical range, and trend pattern), adding these statistical characteristics as supplementary information to the theoretical physical characteristic representation. Alternatively, feature engineering techniques can be used to mine empirical feature patterns related to the missing dimension from the actual observation feature set, and these patterns can be structured and added to the theoretical physical characteristic representation to form a more complete feature description.

[0122] Based on the calibrated and supplemented feature descriptions, consistency checks and conflict resolution are performed on descriptions of the same fault mode originating from different orbits. Irreconcilable feature conflicts are eliminated, resulting in the verified fault modes and calibrated feature descriptions. During this process, for the same fault mode, the system compares its calibrated and supplemented theoretical physical feature description (physical orbit) with the empirical feature description (knowledge orbit) in the context of historical cases. By setting thresholds, numerical features are compared for differences, and discrete features are logically judged to identify feature items with excessive differences or logical contradictions. Conflicts that cannot be reconciled by preset rules (e.g., expert experience, physical laws) are marked as irreconcilable and eliminated. Another approach is to use conflict detection algorithms in multimodal fusion, such as those based on Dempster-Shafer Theory or fuzzy logic, to quantify the degree of conflict between feature descriptions of different orbits. For feature items with a conflict degree exceeding a preset threshold, they are processed according to predefined conflict resolution strategies (e.g., majority rule, expert priority, latest data priority, etc.). If effective resolution is not possible, they are eliminated. After consistency verification and conflict resolution, the system retains those failure modes that have no significant conflicts or whose conflicts have been successfully resolved, and stores their characteristic descriptions in a structured manner.

[0123] The system integrates all validated fault modes and their calibrated feature descriptions to form the fault hypothesis space to be validated. It aggregates all valid fault modes and their corresponding calibrated and supplemented feature descriptions into a structured list or database. Each entry includes a fault mode identifier, its detailed feature description (including physical and empirical dimensions), and a possible confidence score. Alternatively, the system can construct a fault hypothesis graph, where nodes represent validated fault modes, and the node attributes contain their calibrated feature descriptions.

[0124] Through the above technical solutions, this application effectively addresses the problems of statistical bias, missing feature dimensions, and conflicting descriptions of different orbits between theoretical physical feature representations and actual observation feature sets. By acquiring both theoretical physical feature representations and actual observation feature sets, the comprehensiveness of fault mode descriptions is ensured, avoiding the one-sidedness that may result from a single data source. Through statistical difference analysis, parameter range calibration, and supplementation of missing feature dimensions, the feature descriptions output by the theoretical model can more accurately reflect the actual operating status of the equipment, bridging the gap between the model and reality and enhancing the completeness of the feature descriptions. Through consistency verification and conflict resolution, contradictory information from the physical model and knowledge graph can be identified and processed, and unreliable conflicting items can be eliminated, thereby ensuring the internal logical consistency and high reliability of the generated fault hypothesis space to be verified. This fusion and complementary correction mechanism allows the subsequent inference framework to be constructed based on a more accurate, comprehensive, and consistent fault hypothesis space, improving the credibility, interpretability, and overall robustness of fault diagnosis.

[0125] In some of the embodiments described above in this application, it is proposed to calibrate the parameter range of theoretical physical characteristics based on the analysis results and to supplement missing feature dimensions in reverse. However, in the process of implementation, the calibration may lack specific quantitative basis, resulting in inaccurate adjustment. When supplementing missing dimensions, it fails to efficiently extract effective patterns from historical data, making the feature description inaccurate and incomplete, which affects the reliability of subsequent fault hypothesis verification.

[0126] To address this, this application further proposes a method for calibrating the parameter range of theoretical physical characteristics based on the analysis results, and for retrospectively supplementing missing feature dimensions from the actual observation feature set to obtain calibrated and supplemented feature descriptions. This process includes: parsing the analysis results to obtain statistical measures of quantified parameter deviations and labels identifying missing feature dimensions; adaptively adjusting the theoretical range of corresponding physical parameters in the theoretical physical characteristics based on the statistical measures of quantified parameter deviations to obtain the calibrated parameter range; retrieving historical observation data for the corresponding dimension from the actual observation feature set based on the labels identifying missing feature dimensions, and extracting empirical feature patterns characterizing the dimensions; and structurally encapsulating the calibrated parameter range and empirical feature patterns to obtain the calibrated and supplemented feature descriptions.

[0127] Specifically, the steps of analyzing the results to obtain statistical measures of quantified parameter deviations and labels for missing feature dimensions aim to extract quantifiable information from prior analyses (e.g., statistical difference analysis between theoretical physical characteristics and actual observed feature sets) to provide precise guidance for subsequent parameter range calibration and supplementation of missing feature dimensions. For example, parameter deviations can be quantified by calculating statistical indicators such as mean squared error (MSE), mean absolute error (MAE), standard deviation, and confidence intervals, or abnormal fluctuations and trends in parameters can be identified using statistical process control (SPC) charts. Simultaneously, by comparing the dimensional lists of theoretical feature descriptions with those of actual observed feature sets, feature dimensions not included in the theoretical description but present in actual observations can be identified, and unique identifiers or labels can be generated for them. For example, if the theoretical model does not consider the vibration spectrum characteristics of the equipment during operation, but this information is included in the actual observation data, then "vibration spectrum" can be identified as a missing dimension.

[0128] In the step of adaptively adjusting the theoretical range of corresponding physical parameters in the theoretical physical characteristics based on statistical measures of quantified parameter deviations to obtain the calibrated parameter range, this step utilizes quantified deviation information to dynamically correct the range of physical parameters predicted by the theoretical model, making it closer to actual operating conditions and improving the accuracy of theoretical predictions. For example, a statistical method can be used; if statistical measures show that theoretical values ​​are generally too high, the upper and lower limits of the theoretical range can be shifted downwards by a value related to the deviation. Alternatively, the theoretical range can be expanded or contracted based on the standard deviation of the deviation to include the 95% confidence interval of the actual observed values. Furthermore, a regression model or adaptive algorithm such as Kalman filtering can be constructed, taking the statistical measures of quantified parameter deviations as input and outputting adjustment factors or new upper and lower limits for the theoretical range. This model can be trained using historical data to learn how to adaptively adjust according to different deviation patterns.

[0129] In the step of retrieving historical observation data for a missing dimension based on identified feature labels, and extracting empirical feature patterns representing the dimension's characteristics from the actual observation feature set, this step aims to extract valuable, representative regular information from rich historical observation data to supplement and improve the feature description for feature dimensions not covered by theoretical models. For example, for an identified missing dimension (such as vibration spectrum), all relevant vibration spectrum data can be retrieved from historical observation data, and then signal processing techniques (such as Fourier transform and wavelet analysis) can be used to extract features such as dominant frequency, harmonic components, and energy distribution to form empirical feature patterns. Alternatively, clustering algorithms (such as K-means and DBSCAN) can be used to cluster the retrieved historical observation data, identify several common patterns or states under that dimension, and use the statistical characteristics of these patterns (such as mean, variance, and typical curves) to characterize the dimension's characteristics.

[0130] In the step of structurally encapsulating the calibrated parameter value range and empirical feature patterns to obtain the calibrated and supplemented feature description, this step uniformly organizes and formats the calibrated and supplemented feature information to form a complete, consistent, and easily processed feature description, providing reliable input for fault diagnosis. For example, semi-structured data formats such as JSON and XML can be used, or a specific data object / class can be defined to encapsulate the calibrated parameter value range (e.g., {"temperature_range":[min_val,max_val]}) and empirical feature patterns (e.g., {"vibration_pattern":{"dominant_freq":50,"harmonic_ratio":0.2}}) together. Furthermore, this information can also be transformed into concepts and attributes in ontology, or nodes and relationships in a knowledge graph, to facilitate machine understanding and reasoning.

[0131] Through the above technical solution, this application can obtain statistical measures of quantitative parameter deviation and labels for missing feature dimensions by analyzing the results. This provides a data-driven and specific basis for subsequent calibration and supplementation, thereby avoiding blind adjustments and ensuring operational accuracy. Simultaneously, based on the statistical measures of quantitative parameter deviation, the theoretical value range of the corresponding physical parameters in the theoretical physical characteristic performance is adaptively adjusted, allowing the value range to dynamically fit the actual deviation data, improving calibration accuracy. Furthermore, by labeling the missing feature dimensions, historical observation data for the corresponding dimensions is retrieved from the actual observation feature set, and empirical feature patterns characterizing these dimensions are extracted. Effective information from historical data is utilized in a targeted manner to efficiently supplement missing dimensions and enhance feature completeness. The calibrated parameter value range and empirical feature patterns are structurally encapsulated, integrating the calibration and supplementation results to form a unified and operable feature description. This provides a more accurate, complete, and reliable foundation for subsequent fault hypothesis verification, effectively solving the problem of accuracy and completeness in feature description calibration and supplementation.

[0132] In some of the schemes described above in this application, features from different orbits are fused based on calibrated and supplemented feature descriptions to generate a fault hypothesis space to be verified. However, in the implementation process, feature conflicts may lead to unreliable fault hypotheses, thereby reducing the credibility and interpretability of the diagnosis.

[0133] To address this, this application further proposes a method based on calibrated and supplemented feature descriptions to perform consistency verification and conflict resolution on descriptions of the same fault mode originating from different orbits, eliminating irreconcilable feature conflicts, and obtaining the verified fault modes and calibrated feature descriptions. This method includes: separating a theoretical feature subset from the physical orbit and an empirical feature subset from the knowledge orbit from the calibrated and supplemented feature descriptions; performing a numerical and logical comparison of each feature dimension between the theoretical and empirical feature subsets to determine the degree of difference in each feature dimension, and identifying feature dimensions with a degree of difference exceeding a preset sensitivity threshold as potential conflicting feature items; and processing each potential conflicting feature item based on a predefined conflict resolution strategy: if the degree of difference corresponding to the potential conflicting feature item is lower than the reconcilable threshold, then a feature fusion algorithm is applied to generate a reconciled feature value for that feature dimension; if the degree of difference corresponding to the potential conflicting feature item is higher than the reconcilable threshold, then it is determined that the fault mode to which the potential conflicting feature item belongs has an irreconcilable conflict. Remove all fault modes that are determined to have irreconcilable conflicts. For the remaining fault modes, update the original feature description with the generated harmonic feature value to form the fault mode that has passed the verification and the calibrated feature description.

[0134] Specifically, from the calibrated and supplemented feature descriptions, a subset of theoretical features from the physical trajectory and a subset of empirical features from the knowledge trajectory are separated, aiming to clearly distinguish feature information from different sources. This can be achieved by attaching a source label (e.g., "physical model deduction" or "knowledge graph extraction") to each feature element during feature description generation, using label filtering. Alternatively, the system can parse features based on their semantic content and structure; for example, numerical features from digital twin simulation results can be classified as theoretical features, while descriptive features from historical failure cases, expert rules, or ontology definitions can be classified as empirical features. This separation ensures the relevance and accuracy of subsequent comparisons, avoiding confusion between information from different sources before comparison.

[0135] The theoretical feature subset and the empirical feature subset are compared numerically and logically along each feature dimension to determine the degree of difference in each feature dimension. Feature dimensions with a difference exceeding a preset sensitivity threshold are marked as potentially conflicting feature items. For numerical features, the degree of difference can be quantified by calculating the absolute difference, relative percentage difference, or statistical distance (such as Euclidean distance). For example, if the theoretically predicted temperature is 100℃, while empirical data indicates 120℃, the degree of difference is 20℃ or 20%. For logical or categorical features, the degree of difference can be determined by Boolean comparison (whether they are consistent) or semantic similarity calculation. When the degree of difference in any feature dimension exceeds the preset sensitivity threshold, that feature dimension is marked as a potentially conflicting feature item. This threshold can be configured based on equipment type, fault sensitivity, or domain expert experience; for example, it can be set to a numerical deviation exceeding 5% or a logical inconsistency.

[0136] Based on this, each potential conflict feature is processed according to a predefined conflict resolution strategy. If the degree of difference corresponding to the potential conflict feature is lower than the reconciliation threshold, a feature fusion algorithm is applied to generate a reconciled feature value for that feature dimension. For example, when there is a small deviation between theoretical and empirical values, algorithms such as weighted averaging, Bayesian fusion, or fuzzy logic fusion can be used to combine the information from both to generate a more robust reconciliation value. The weight allocation can be based on the reliability of the data source, update frequency, or historical performance. If the degree of difference corresponding to the potential conflict feature is higher than the reconciliation threshold, the failure mode to which the potential conflict feature belongs is determined to have an irreconcilable conflict. This means that there is a fundamental discrepancy between the theoretical model and empirical knowledge on key features, which cannot be resolved by simple fusion. For example, if the theoretical model predicts an increase in vibration frequency, while empirical knowledge indicates a decrease in vibration frequency, and the difference between the two is huge, it is considered an irreconcilable conflict.

[0137] All fault modes identified as having irreconcilable conflicts are removed. For the remaining fault modes, the original feature descriptions are updated using the generated harmonic feature values, forming the validated fault mode and its calibrated feature description. Fault modes identified as having irreconcilable conflicts are removed from the fault hypothesis space to avoid introducing uncertainty or incorrect diagnostic directions. For validated fault modes, all harmonic feature dimensions in their original feature descriptions are updated with new harmonic feature values. This ensures that the fault hypothesis space contains only fault modes that have high consistency between the physical model and the domain knowledge graph or have successfully harmonized conflicts, thus providing a reliable foundation for subsequent evidence gathering and reasoning.

[0138] Through the above technical solution, this application introduces a systematic conflict handling mechanism, solving the problem of unreliable fault hypotheses caused by feature conflicts during feature fusion, and ensuring that the generated fault hypothesis space is more reliable and consistent. Specifically, by separating the theoretical feature subset from the physical orbit and the empirical feature subset from the knowledge orbit, conflicts can be identified and handled in a targeted manner, avoiding error propagation and laying the foundation for subsequent conflict detection. By comparing the numerical and logical values ​​of each feature dimension, the differences in each dimension are accurately identified, rather than making an overall judgment, thus improving the accuracy of conflict localization. By marking feature dimensions with differences exceeding a preset sensitivity threshold as potential conflict feature items, the degree of conflict is quantified, ensuring that only significant differences are processed and avoiding resource waste. Each potential conflict feature item is processed based on a predefined conflict resolution strategy, providing a structured method to deal with conflicts. If the difference is lower than the reconciliation threshold, a feature fusion algorithm is applied to generate reconciled feature values, achieving conflict reconciliation and retaining useful information. If the difference is higher than the reconciliation threshold, the fault mode is determined to have irreconcilable conflicts, and unreliable items are actively excluded. All fault modes determined to have irreconcilable conflicts are removed, directly eliminating high-risk hypotheses and improving overall reliability. For the remaining fault modes, the generated harmonic eigenvalues ​​are used to update the original feature descriptions, ensuring feature data consistency. This results in validated fault modes and calibrated feature descriptions, providing reliable input for diagnosis and improving the quality of the fault hypothesis space to be validated, as well as the credibility and interpretability of subsequent diagnoses.

[0139] In some of the solutions described above in this application, a multimodal evidence type is assigned to each fault hypothesis in the fault hypothesis space to be verified in order to define the evidence requirements. However, in the process of implementation, the selection of evidence type may fail to effectively integrate the theoretical physical characteristics and historical case context, resulting in inaccurate evidence collection, failure to ensure physical consistency and empirical reliability, and thus affecting the credibility of the diagnostic conclusion.

[0140] To address this, this application further proposes assigning the required multimodal evidence types to each fault hypothesis in the fault hypothesis space to be verified. This method extracts the theoretical physical characteristics and historical case context associated with each fault hypothesis in the fault hypothesis space. This step aims to lay the foundation for subsequent evidence type assignment, ensuring that the assignment process simultaneously considers the theoretical basis of the physical model derivation and the practical support of historical experience. Extracting theoretical physical characteristics refers to the changes in physical quantities and signal characteristics that the equipment should exhibit under a specific fault mode, derived from the equipment's physical model, such as abnormal vibration spectrum, temperature rise, and current fluctuations. These characteristics directly reflect the fault mechanism. Extracting historical case context refers to historical fault records, maintenance logs, expert experience, etc., related to the current fault hypothesis, obtained from the domain knowledge graph. This may include the evidence types, observation methods, or feature patterns relied upon for successfully diagnosing this type of fault in the past. For example, theoretical physical characteristics of bearing wear failures may include an increase in vibration amplitude at a specific frequency, while historical case context may indicate that such failures are typically verified by collecting vibration data with an accelerometer and combining it with infrared thermal imager observation of bearing temperature.

[0141] Based on the theoretical physical characteristics, the key physical feature dimensions necessary for verifying each fault hypothesis are identified. The purpose of this step is to select the most critical and distinguishable observation dimensions from the theoretical physical characteristics, thereby avoiding redundant data collection and improving the efficiency and relevance of evidence gathering. Key physical feature dimensions refer to physical quantities or their trends that play a decisive role in fault diagnosis. For example, for imbalance faults in rotating machinery, the key physical feature dimension might be the vibration amplitude of the rotational frequency and its harmonics. For insulation aging faults in electrical equipment, it might be the intensity of the partial discharge signal or the dielectric loss angle. The identification process can be based on physical model sensitivity analysis to determine which physical quantities are most sensitive to specific fault modes, or based on expert experience rules to directly specify key observation points.

[0142] Based on pre-defined mapping rules between feature dimensions and multimodal sensing capabilities, the key physical feature dimension is transformed into an initial set of candidate evidence types. This step associates the abstract key physical feature dimension with specific sensing technology capabilities, thereby generating selectable preliminary evidence types. The pre-defined mapping rules between feature dimensions and multimodal sensing capabilities are a pre-established knowledge base or rule set that defines which types of sensors or detection devices can effectively observe different physical features (such as vibration, temperature, acoustics, electricity, images, etc.). For example, vibration features can be mapped to accelerometers and displacement sensors. Temperature features can be mapped to thermocouples and infrared thermal imagers. Acoustic features can be mapped to microphones and ultrasonic sensors. Image features can be mapped to visible light cameras and industrial endoscopes. Through these rules, the system can transform the need to "observe vibration" into evidence types that "require the collection of vibration data."

[0143] Based on this, an empirical selection process is performed on the initial set of candidate evidence types, using the valid evidence types recorded in the context of the historical case. This step aims to incorporate historical experience to optimize and filter the theoretically derived initial candidate evidence types, thereby improving the practicality and effectiveness of evidence collection. Valid evidence types recorded in the context of historical cases refer to evidence types that have proven to be effective, reliable, and highly discriminative in the past successful diagnosis of similar faults. For example, if historical records show that a certain fault is typically located quickly on-site by an abnormal temperature reading from an infrared thermal imager, then even if vibration analysis could theoretically provide information, infrared thermal imager data, as a valid evidence type, would receive a higher weight in the selection process. Empirical selection can be achieved by ranking and filtering different evidence types in historical cases based on indicators such as diagnostic success rate, collection cost, and ease of use, or by using an expert system to directly exclude or recommend certain evidence types based on historical experience rules.

[0144] Based on the physical consistency constraints and spatiotemporal acquisition constraints defined in the inference framework, feasibility verification is performed on the empirically optimized evidence types to obtain the multimodal evidence types required for each fault hypothesis. This step is crucial in determining the evidence types, ensuring that the selected evidence types are not only theoretically and empirically valid but also practically feasible and consistent with the overall inference logic. Physical consistency constraints refer to the predefined restrictions in the inference framework regarding the relationships between different physical quantities. For example, in certain fault modes, abnormal vibration is necessarily accompanied by a temperature increase, or certain physical quantities should be maintained within specific ranges under specific operating conditions. Spatiotemporal acquisition constraints refer to the limitations on the time, spatial range, frequency, etc., of evidence acquisition. For example, some sensors can only be installed in specific locations, or some data can only be collected when the equipment is shut down. Feasibility verification checks whether the empirically optimized evidence types can meet these constraints. For example, if a certain evidence type requires the acquisition of internal equipment data, but the equipment cannot be shut down, then that evidence type may be deemed infeasible. The verification process may include simulating the acquisition environment, checking sensor availability, and evaluating acquisition costs and time.

[0145] Through the above technical solution, this application systematically combines theoretical physical characteristics with historical case context to optimize the evidence type assignment process, ensuring accurate and reliable evidence collection and thus improving the credibility of fault diagnosis. Specifically, for each fault hypothesis in the fault hypothesis space to be verified, the associated theoretical physical characteristics and historical case context are extracted. This step ensures that evidence type assignment is based on physical model deduction and historical knowledge, avoiding evidence floating or deviating from the physical path, and providing a dual-track foundation for subsequent operations. Based on the theoretical physical characteristics, the key physical feature dimensions necessary for verifying each fault hypothesis are identified, directly determining the core verification points from the physical model, ensuring that evidence collection focuses on necessary features rather than redundant dimensions, and enhancing targeting. According to the preset mapping rules between feature dimensions and multimodal sensing capabilities, the key physical feature dimensions are converted into an initial candidate evidence type set. The rule-based mapping transforms physical requirements into specific sensing options, providing a preliminary evidence type framework and avoiding defocus caused by arbitrary selection. Based on the valid evidence types recorded in the historical case context, the initial candidate evidence type set is empirically optimized, and the candidate set is optimized by referring to historical valid data, incorporating empirical knowledge to improve the reliability and adaptability of evidence. Based on the physical consistency constraints and spatiotemporal acquisition constraints defined in the reasoning framework, the feasibility of empirically selected evidence types is verified. The actual feasibility of evidence types is verified by combining physical and spatiotemporal constraints, ensuring that the acquisition task is executable and conforms to the overall reasoning logic. This yields the multimodal evidence types required for each fault hypothesis, achieving efficient and accurate evidence assignment.

[0146] In some of the solutions described above in this application, a reasoning skeleton is generated through dual-track coupling and constraint injection to define the fault hypothesis space to be verified, multimodal evidence requirements, and embedded physical and spatiotemporal constraints. However, in this process, due to the lack of an effective structured organization mechanism, physical consistency constraints and spatiotemporal acquisition constraints may be disorderly associated with fault hypotheses and evidence types, resulting in scattered and chaotic verification rules and acquisition tasks. This makes it difficult to ensure the logical consistency and traceability of the reasoning skeleton, thereby affecting the collaborative efficiency of subsequent evidence acquisition and the accuracy of the reasoning process.

[0147] To address this, this application further proposes a method for assembling a reasoning skeleton by associating physical consistency constraints and spatiotemporal acquisition constraints as structured metadata, respectively, with the latter linked to the corresponding fault hypotheses and multimodal evidence types in the fault hypothesis space to be verified. Specifically, this method includes: creating a unique exploration step identifier for each fault hypothesis in the fault hypothesis space to be verified; associating the physical consistency constraints used to verify each fault hypothesis as verification rule metadata for each fault hypothesis with the corresponding exploration step identifier; associating the multimodal evidence type required to verify each fault hypothesis and the corresponding spatiotemporal acquisition constraints as acquisition task metadata for each fault hypothesis with the corresponding exploration step identifier; performing topological sorting of all exploration step identifiers based on the fault propagation topology contained in the associated knowledge subgraph path; and structurally encapsulating each exploration step identifier and its associated verification rule metadata and acquisition task metadata according to the topological sorting order to obtain the reasoning skeleton.

[0148] This involves creating a unique exploration step identifier for each fault hypothesis in the fault hypothesis space to be verified. The aim is to assign a unique identifier to the verification process of each fault hypothesis. This identifier serves to individually manage and track the verification status, related constraints, and evidence requirements of each fault hypothesis, ensuring that the exploration activities of each hypothesis can be clearly identified and located in complex reasoning processes, avoiding information confusion. For example, a globally unique identifier (UUID) or a hash value based on the fault hypothesis content (such as fault mode, involved component ID, etc.) can be used as the exploration step identifier to ensure its uniqueness throughout the system. Alternatively, sequentially increasing integers or string codes, combined with the batch number of the current fault diagnosis task, can be used to form a hierarchical identifier for easy management and querying.

[0149] Using the physical consistency constraints used to verify each fault hypothesis as verification rule metadata for each fault hypothesis, and associating them with the corresponding exploration step identifier, means encapsulating the physical principles and logical relationships related to a specific fault hypothesis in the form of structured data. This metadata is directly bound to the corresponding exploration step identifier, clearly defining the physical laws and logical rules that must be followed when verifying the fault hypothesis. For example, verification rule metadata can be stored in JSON or XML format, containing the type of constraint (such as energy conservation, mass conservation, kinematic relationships, thermodynamic equilibrium, etc.), the physical quantities involved, the allowed threshold range, and the specific verification logic expression. Alternatively, physical consistency constraints can be defined as object instances with specific attributes and methods in an object-oriented manner, and then serialized and stored as metadata for easy programmatic invocation and execution.

[0150] Encapsulating the types of multimodal evidence required to verify each fault hypothesis, along with the corresponding spatiotemporal acquisition constraints, as metadata for each fault hypothesis's acquisition task, and linking it to the corresponding exploration step identifier, means encapsulating the types of multimodal evidence needed to verify a specific fault hypothesis and the spatiotemporal conditions for acquiring this evidence in the form of structured data. This metadata is also linked to the corresponding exploration step identifier, providing clear instructions to the subsequent evidence acquisition agent, guiding it on how to efficiently and accurately acquire the required evidence. For example, the acquisition task metadata can include information such as evidence modality type (e.g., vibration signal, thermal imaging image, acoustic data, current waveform, etc.), the spatial coordinate range of the target device component, the recommended acquisition time window, the acquisition frequency, and the required sensor type and accuracy requirements, organized in key-value pairs or structure arrays. Alternatively, the acquisition task metadata can be designed as an executable script or configuration instruction set, which the agent can directly parse and execute to complete evidence acquisition, thereby achieving automation and intelligence.

[0151] Topological sorting of all investigation step identifiers, based on the fault propagation topology contained in the associated knowledge subgraph paths, refers to arranging the investigation steps of all fault hypotheses to be verified in an ordered manner according to the inherent logical relationship of fault occurrence and propagation in the device. Topological sorting is a method for linearly sorting the vertices of a directed acyclic graph (DAG), such that for any directed edge from vertex u to vertex v, u always appears before v. Here, by performing topological sorting on the investigation step identifiers, the aim is to ensure that the verification order of fault hypotheses conforms to the actual fault occurrence and propagation logic, avoiding reverse reasoning or omitting key intermediate steps, thereby improving the efficiency and accuracy of reasoning. For example, the Kahn algorithm can be used, maintaining a queue of vertices with an in-degree of 0, gradually removing vertices and updating the in-degree of their adjacent vertices until all vertices are sorted. Alternatively, a depth-first search (DFS) algorithm can be used, adding vertices to the result list during backtracking to obtain a reverse topological sort, which can then be reversed to obtain a forward topological sort.

[0152] Following the topological sorting order, each exploration step identifier and its associated verification rule metadata and data acquisition task metadata are structurally encapsulated to obtain the inference skeleton. This means integrating the topologically sorted exploration step identifiers, along with their respective associated verification rule metadata and data acquisition task metadata, into a unified data structure with clear hierarchy and logical relationships. This encapsulated data structure is the inference skeleton, which provides a unified and standardized input for subsequent collaborative evidence collection and traceable inference, ensuring the orderliness and controllability of the entire diagnostic process. For example, the inference skeleton can be encapsulated as a directed acyclic graph (DAG) data structure, where nodes represent exploration step identifiers, node attributes contain verification rule metadata and data acquisition task metadata, and edges represent the dependencies of fault propagation. Alternatively, it can be encapsulated as an ordered list or array, where each element is a structure or object containing exploration step identifiers, verification rule metadata, and data acquisition task metadata, and the order of the list is the result of the topological sorting.

[0153] Through the above technical solutions, this application effectively solves the problem of disorder in the generation of the reasoning skeleton, and improves the logical consistency and traceability of the diagnostic process. Specifically, by creating a unique exploration step identifier for each fault hypothesis, fine-grained management of the fault hypothesis verification process is achieved, avoiding confusion and omissions in verification tasks. Physical consistency constraints are directly associated with the exploration step identifier as verification rule metadata, ensuring a tight binding between physical constraints and specific fault hypotheses, effectively preventing "physical decoupling" and making the verification process more rigorous and reliable. Simultaneously, multimodal evidence types and spatiotemporal acquisition constraints are associated with the exploration step identifier as acquisition task metadata, clarifying the specific needs and scope of evidence acquisition, greatly improving the targeting and efficiency of evidence acquisition. Furthermore, based on the fault propagation topology contained in the associated knowledge subgraph path, all exploration step identifiers are topologically sorted, ensuring that the verification order of fault hypotheses strictly follows the actual propagation logic of the fault, avoiding the "logical defocusing" problem and ensuring the rationality of the reasoning path. By structurally encapsulating the identification of exploration steps and associated metadata according to the topological sorting order, a clear, traceable and logically rigorous reasoning framework is formed, providing clear instructions and basis for the subsequent collaborative evidence collection of heterogeneous intelligent agents, thereby improving the accuracy, reliability and efficiency of the entire fault diagnosis process.

[0154] In some of the solutions mentioned above in this application, it is proposed to schedule multiple heterogeneous intelligent agents to perform collaborative evidence collection in order to collect multimodal evidence. However, in the process of implementation, due to the diversity of heterogeneous intelligent agents and the complexity of tasks, evidence collection may lack atomic task description, agent allocation may be mismatched, constraint transformation may be inaccurate, and timing coordination may be improper, resulting in spatiotemporal asynchrony of evidence data and difficulty in conflict detection, thereby affecting the accuracy and efficiency of subsequent fusion and reasoning.

[0155] In response, this application further proposes a method for scheduling multiple heterogeneous intelligent agents to perform collaborative evidence collection based on a reasoning skeleton, see [link to relevant documentation]. Figure 5 The method includes the following steps: 501. Analyze the reasoning framework, extract the fault hypothesis to be verified, and the associated multimodal evidence requirements, physical constraints, and spatiotemporal acquisition constraints.

[0156] This step aims to extract all necessary information from the pre-built reasoning skeleton to guide the subsequent evidence collection process, ensuring the relevance and effectiveness of the collection. Specifically, a parser based on structured data formats (such as XML and JSON) can be used to parse the reasoning skeleton layer by layer, identifying and extracting predefined fields to obtain information such as fault hypotheses, multimodal evidence requirements, physical constraints, and spatiotemporal collection constraints. Alternatively, Natural Language Processing (NLP) techniques can be combined. If the reasoning skeleton contains some unstructured or semi-structured descriptions, semantic analysis, entity recognition, and relation extraction techniques can be used to intelligently identify and extract the required information from the text descriptions.

[0157] 502. Based on the multimodal evidence requirement and the spatiotemporal acquisition constraint, generate an atomic evidence acquisition task description for each fault hypothesis, and assign a matching heterogeneous intelligent agent to execute each atomic evidence acquisition task description.

[0158] This step decomposes complex evidence collection tasks into smaller, more manageable atomic tasks, ensuring that each task is executed by the most suitable heterogeneous agent, thereby optimizing resource utilization and improving task execution efficiency. Specifically, based on a pre-defined rule base, multimodal evidence requirements (e.g., vibration data, thermal imaging, acoustic signals, etc.) and spatiotemporal acquisition constraints (e.g., specific spatial locations, time windows) are combined to generate atomic evidence collection task descriptions containing task identifiers, evidence modality types, spatial coordinates, acquisition time windows, and associated fault hypothesis identifiers. Simultaneously, the system maintains a heterogeneous agent capability registry, recording the core evidence modality types supported by each agent and its current schedulable state. Tasks are assigned to the most suitable online agent using a matching algorithm (e.g., priority-based greedy matching or cost-based optimal matching). Alternatively, an ontology encompassing concepts such as evidence modalities, agent capabilities, and device components is constructed. An ontology reasoning mechanism is used to automatically generate atomic evidence collection task descriptions, and an agent allocation is performed through a semantic matching engine.

[0159] 503. Convert the physical constraints and the spatiotemporal acquisition constraints into acquisition command parameters that can be executed by the heterogeneous intelligent agent, and coordinate the task startup sequence of the heterogeneous intelligent agent based on the logical relationship between the steps in the inference skeleton.

[0160] This step aims to transform high-level physical and spatiotemporal constraints into specific operational instructions that heterogeneous agents can directly understand and execute, and to orchestrate the execution order of these agents to achieve efficient and synchronous data acquisition. Specifically, a standardized set of API interfaces or instruction sets can be predefined for different types of heterogeneous agents, mapping physical constraints (e.g., sampling frequency, sensor range) and spatiotemporal acquisition constraints (e.g., acquisition duration, triggering conditions) to parameters of these API interfaces. Simultaneously, based on the fault propagation topology or dependencies between exploration steps defined in the inference framework, a task dependency graph is constructed. Topology sorting or event-driven mechanisms are used to coordinate the startup sequence of heterogeneous agents, ensuring that subsequent tasks are started only after the preceding tasks are completed. Furthermore, a dedicated adapter can be developed for each heterogeneous agent, responsible for converting general constraints into a specific instruction format that the agent can understand. A central scheduler then generates a global task execution plan based on the logical relationships in the inference framework.

[0161] 504. Schedule the multiple heterogeneous intelligent agents to execute atomized evidence collection tasks according to the collection command parameters, and make the multimodal raw evidence data output by the heterogeneous intelligent agents accompanied by a unified spatiotemporal synchronization mark.

[0162] This step is crucial for the actual data acquisition process, ensuring that all acquired data carries precise timestamps and spatial references, laying the foundation for subsequent data fusion and analysis. Specifically, commands containing acquisition instructions can be sent to assigned heterogeneous agents via message queues or publish / subscribe mechanisms. Each heterogeneous agent, while acquiring data, uses a Network Time Protocol (NTP) server or a built-in GPS module to obtain high-precision timestamps and spatial coordinates, embedding this information as a unified spatiotemporal synchronization marker into its output multimodal raw evidence data. Alternatively, the heterogeneous agents can stream the acquired raw data to a central data recorder, which uses a hardware synchronization module (e.g., PTP protocol) or software algorithm to spatiotemporally synchronize data across different modalities, then adds a unified spatiotemporal synchronization marker to each data record.

[0163] Through the above technical solutions, this application effectively addresses issues such as missing atomized task descriptions, mismatched agent allocation, inaccurate constraint transformation, and improper timing coordination in the evidence collection process. Specifically, by accurately analyzing the inference framework, it can extract the fault hypotheses to be verified, multimodal evidence requirements, physical constraints, and spatiotemporal collection constraints, providing clear and comprehensive guidance for subsequent evidence collection and avoiding blind or redundant collection actions. Based on this, an atomized evidence collection task description is generated for each fault hypothesis, and matching heterogeneous agents are assigned, making the collection tasks granular and manageable, and fully utilizing the professional capabilities of heterogeneous agents to improve the adaptability and efficiency of task execution. Simultaneously, physical constraints and spatiotemporal collection constraints are converted into collection instruction parameters executable by heterogeneous agents, ensuring the accuracy and compliance of the collection process. Furthermore, based on the logical relationships between steps in the inference framework, the task startup sequence of heterogeneous agents is coordinated, optimizing the task execution process, avoiding resource conflicts and waiting, and improving overall collection efficiency. By scheduling heterogeneous intelligent agents to execute atomic tasks and attaching unified spatiotemporal synchronization markers to the output multimodal raw evidence data, the precise alignment of all collected evidence data in the time and space dimensions is ensured. This greatly facilitates subsequent conflict detection, cross-validation, and multimodal fusion analysis, thereby improving the accuracy and reliability of fault diagnosis.

[0164] In some of the embodiments described above in this application, an evidence collection task description is generated for each fault hypothesis and an agent is assigned to achieve collaborative evidence collection. However, in this process, the task description may not be accurate enough, and the assignment may not match the capabilities of the agent, resulting in low evidence collection efficiency, wasted resources, or insufficient data to accurately verify the fault hypothesis.

[0165] In response, this application further proposes a method based on the multimodal evidence requirement and the spatiotemporal acquisition constraint, which generates an atomic evidence acquisition task description for each fault hypothesis, and assigns a matching heterogeneous intelligent agent to execute each atomic evidence acquisition task description.

[0166] Specifically, the method includes the following steps: This step involves parsing the required evidence modality from the multimodal evidence requirement and extracting the target spatial coordinates and acquisition time window from the spatiotemporal acquisition constraints. The aim is to precisely define the required evidence type, acquisition location, and time frame, providing clear guidance for subsequent evidence acquisition tasks. For example, semantic analysis or pattern matching of the multimodal evidence requirement can extract the required evidence modality type, such as vibration signals, thermal imaging images, acoustic data, or current and voltage waveforms, from the structured or semi-structured requirement description. Simultaneously, predefined geographical coordinates, equipment component numbers, timestamp ranges, or time period descriptions in the spatiotemporal acquisition constraints can be parsed and converted into operable target spatial coordinates (such as 3D coordinates, area identifiers) and acquisition time windows (such as start time, end time, and duration).

[0167] Based on the parsed evidence modality type, target spatial coordinates, and acquisition time window, an atomic evidence acquisition task description is constructed. This atomic task description encapsulates a task identifier, evidence modality type, spatial coordinates, acquisition time window, and associated fault hypothesis identifier. This step integrates all necessary information into a single, unambiguous task unit, facilitating subsequent scheduling and execution. For example, the parsed information can be encapsulated into a structured data object containing a unique task identifier, the required evidence modality type, precise target spatial coordinates, a predetermined acquisition time window, and a unique identifier for the fault hypothesis associated with the task. Alternatively, this information can be transformed into a set of standardized API call parameters, which define all the inputs required for the agent to execute the acquisition task.

[0168] Access a predefined heterogeneous agent capability registry to retrieve the core evidence modalities supported by each heterogeneous agent and its current schedulable status. This step aims to obtain real-time information on the capabilities and availability of various heterogeneous agents, providing a basis for intelligent task allocation. For example, this can be achieved by sending a query request to a centralized agent management service that maintains a database recording the IDs of all registered agents, the types of sensors they support (i.e., core evidence modalities), processing capabilities, and their current schedulable status (e.g., online / offline, busy / idle). Alternatively, a distributed service discovery mechanism can be used, where agents publish their capabilities and status to the registry upon startup and update it as their status changes; the task allocator can then subscribe to this information or query it on demand.

[0169] Based on the evidence modality type in the atomic evidence collection task description, a matching decision is made based on the core evidence modality type recorded in the heterogeneous agent capability registry and the current schedulable state. A heterogeneous agent that supports the corresponding evidence modality type and is currently online is assigned to each atomic evidence collection task description. This step achieves intelligent matching between tasks and agents, ensuring that tasks can be executed by agents with the corresponding capabilities and currently available. For example, a rule engine can be designed. This engine searches the agent capability registry for all agents that support the evidence modality type specified in the atomic task description and are currently "online" or "idle," and selects the best-matching agent for allocation based on preset priorities (such as closest distance, lowest load, highest accuracy, etc.). Alternatively, an optimization algorithm can be used, such as a greedy algorithm or a multi-objective optimization algorithm, to comprehensively consider multiple factors such as agent capability matching degree, geographical location, current load, and battery life to find the optimal agent allocation scheme for each task.

[0170] Based on the evidence modality type in the description of the atomized evidence collection task, a match is made with the heterogeneous agent capability registry. This step further enhances the accuracy and reliability of the matching process, ensuring the smooth execution of the collection task and the acquisition of high-quality evidence. For example, a capability matrix can be constructed, where rows represent agents, columns represent evidence modality types, and matrix cells indicate whether the agent supports that modality. During assignment, this matrix is ​​directly queried to confirm whether the agent possesses the collection capability for the required modality. Alternatively, more advanced semantic matching techniques can be used, which not only match precise modality type names but also identify agents with similar or compatible capabilities.

[0171] Through the above technical solution, this application effectively solves the problems of inaccurate task descriptions and mismatched agent capabilities leading to low evidence collection efficiency, resource waste, or insufficient data to accurately verify fault hypotheses by precisely generating atomic evidence collection task descriptions and dynamically matching and allocating them based on the agent capability registry. Specifically, by parsing the evidence modality type to be collected from multimodal evidence requirements and parsing the target spatial coordinates and collection time window from spatiotemporal collection constraints, it ensures that each collection task has a clear evidence type, spatial range, and time window, greatly improving the accuracy and operability of the task description and avoiding ambiguity and uncertainty in the collection process. On this basis, an atomic evidence collection task description is constructed, encapsulating the task identifier, evidence modality type, spatial coordinates, collection time window, and associated fault hypothesis identifier, making each task an independent and clearly defined unit. This not only facilitates subsequent scheduling and execution but also effectively reduces conflicts and resource contention between tasks. Furthermore, by accessing the predefined heterogeneous agent capability registry, the core evidence modality types supported by each heterogeneous agent and its current schedulable status are obtained in real time, enabling the system to dynamically grasp the agent's capabilities and availability information. Based on this, a matching decision is made according to the evidence modality type in the atomic evidence collection task description, combined with the core evidence modality type recorded in the registry and the current schedulable status. A heterogeneous agent supporting the corresponding evidence modality type and currently online is assigned to each atomic evidence collection task description. This precise matching mechanism ensures that tasks are assigned to agents that truly possess the corresponding collection capabilities and are currently available, thereby optimizing resource utilization efficiency and avoiding collection failures or low data quality due to capability mismatches. Simultaneously, by continuously matching the evidence modality type in the atomic evidence collection task description with the heterogeneous agent capability registry, the accuracy and reliability of the matching process are further enhanced, ensuring the smooth execution of collection tasks and the acquisition of high-quality evidence. Overall, the above technical solution makes the evidence collection process more intelligent, automated, and efficient, providing high-quality, highly reliable multimodal evidence support for subsequent fault diagnosis and insight, and improving the accuracy and reliability of the entire data processing and insight method.

[0172] In some of the solutions mentioned above in this application, conflict detection and cross-validation are proposed to form a structured evidence network based on real-time acquired evidence during the collection process. However, in this process, without an effective mechanism to standardize the evidence format, detect cross-modal conflicts in real time, perform targeted retesting and verification, and systematically integrate evidence relationships, the evidence data may contradict each other due to modal differences and spatiotemporal inconsistencies, which cannot reliably support subsequent reasoning and reduce the credibility of the diagnostic conclusion.

[0173] To address this, this application proposes a method for performing conflict detection and cross-validation based on real-time acquired evidence during the data collection process, thereby forming a structured evidence network with consistency markers. (See [link to relevant documentation]). Figure 6 The method includes the following steps: 601. Perform modal analysis and spatiotemporal synchronization alignment on the real-time acquired evidence to generate multiple standardized evidence nodes and form a standardized evidence node set. Each standardized evidence node includes evidence content, modal type, spatial location, and acquisition timestamp.

[0174] Specifically, real-time acquired evidence may originate from multiple sensors or data sources, such as vibration sensors, thermal imagers, acoustic sensors, and pressure gauges. Modal analysis aims to identify and extract core information from this raw data, such as extracting frequency features from vibration signals, temperature distribution from thermal images, and anomalous acoustic signatures from acoustic data. Spatiotemporal synchronization and alignment unifies this data from different modalities, which may have been collected at different times or spatial locations, onto a common spatiotemporal reference. For example, high-precision timestamp synchronization technologies (such as NTP or PTP protocols) can ensure temporal consistency of data from different sensors, and geolocation systems (such as GPS) or the device's internal coordinate system can map data from different spatial locations to a unified physical model of the device. Through this step, raw, heterogeneous evidence data is converted into standardized evidence nodes in a unified format, laying the foundation for subsequent conflict detection and reasoning. For example, vibration data can be analyzed as evidence of "high-frequency vibration anomaly," with the modal type being "vibration," the spatial location being "bearing A," and the acquisition timestamp being "T1." Meanwhile, the thermal imaging data was analyzed as evidence of "local overheating", with the modality type being "thermal imaging", the spatial location being "bearing A", and the acquisition timestamp being "T1+Δt".

[0175] 602. Based on the modal type and content of each standardized evidence node in the standardized evidence node set, perform cross-modal logical consistency verification and numerical compatibility analysis to identify and mark conflicting evidence node pairs that contradict each other.

[0176] Specifically, cross-modal logical consistency verification refers to assessing whether there are logical contradictions or irrationalities between different modal evidence. For example, if vibration evidence indicates severe wear in bearing A, while thermal imaging evidence shows that bearing A's temperature is normal, this may constitute a logical inconsistency. This verification can be performed using a predefined expert rule base, ontological reasoning, or machine learning models. Numerical compatibility analysis focuses on quantifying whether the numerical differences between evidence data are within acceptable limits. For example, if two different types of temperature sensors (one infrared, one contact) collect significantly different temperature readings at the same location and time, numerical incompatibility may exist. This analysis can be based on comparisons using statistical methods (such as significance tests) or physical model predictions. Through this step, the system can proactively identify potential contradictions between evidence and mark them as conflicting evidence pairs, providing targets for subsequent verification. For example, if vibration modal evidence indicates "severe bearing wear," while acoustic modal evidence indicates "smooth bearing operation without abnormal noise," these two evidence nodes may be marked as a conflicting evidence pair.

[0177] 603. For the marked conflict evidence node pairs, trigger targeted cross-validation: schedule heterogeneous agents related to the conflict mode, perform retesting based on the spatial location involved in the conflict, generate verification evidence nodes, and add the verification evidence nodes to the standardized evidence node set.

[0178] Specifically, when conflicting evidence nodes are detected, the system does not make a judgment but actively seeks supplementary information to resolve the conflict. Directed cross-validation means that the system intelligently selects the most suitable heterogeneous intelligent agent (e.g., if the conflict involves vibration data, a high-precision vibration sensor or robot may be scheduled for close-range detection; if it involves visual data, a drone may be scheduled for high-resolution image acquisition) to perform a retest task based on the modal type and spatial location of the conflict. The retest results will generate new verification evidence nodes. These nodes are also aligned with modal analysis and spatiotemporal synchronization and added to the standardized evidence node set, thereby enriching the evidence set and providing more evidence for resolving the conflict. For example, regarding the conflict between "severe bearing wear" and "smooth bearing operation," the system can schedule a mobile robot equipped with a high-precision accelerometer to go to the precise location of bearing A to collect secondary vibration data to obtain more reliable vibration evidence.

[0179] 604. Using all nodes in the standardized evidence node set as network nodes, and using the supporting relationships, contradictory relationships, or verification relationships between nodes as edges, construct the structured evidence network.

[0180] Specifically, after all evidence (including original standardized evidence and verification evidence) has been processed, the system organizes these evidence nodes into a graph structure. Nodes in the network represent individual standardized pieces of evidence, while edges represent the relationships between them. Supporting relationships indicate that two pieces of evidence corroborate each other or jointly support a hypothesis. Conflicting relationships indicate that two pieces of evidence contradict each other. Verifying relationships indicate that a piece of evidence was generated to verify another piece of conflicting evidence. These relationships can be automatically identified and established through pre-defined rules, domain knowledge, or machine learning models. Through this step, the originally discrete evidence is integrated into an evidence network with rich semantics and topological structure, including consistency markers (e.g., representing conflict states, verification states, or confidence levels through node or edge attributes), providing a comprehensive and traceable evidentiary foundation for subsequent reasoning and decision-making. For example, in a network, a thermal imaging evidence node for "bearing overheating" can be connected to a knowledge node for "poor lubrication" through a "cause" relationship, and to a vibration evidence node for "normal vibration" through a "contradictory" relationship. Meanwhile, a newly acquired verification evidence node for "high-precision vibration anomaly" may form a "verification" relationship with the original "normal vibration" node and a "support" relationship with the "bearing overheating" node.

[0181] Through the above technical solutions, this application effectively addresses the problems of modal differences, spatiotemporal inconsistencies, and contradictions that may arise during the acquisition of multimodal evidence. By performing modal analysis and spatiotemporal synchronization alignment on real-time acquired evidence, the standardization and unification of multi-source heterogeneous evidence are achieved, providing a solid foundation for subsequent processing and avoiding initial inconsistencies caused by different data formats and spatiotemporal misalignments. By performing cross-modal logical consistency checks and numerical compatibility analysis, the system can proactively and in real-time identify and mark potential conflicts between pieces of evidence, ensuring the timely exposure of contradictory points and improving the internal consistency of the evidence. For identified conflicts, a targeted cross-validation mechanism is triggered, scheduling heterogeneous intelligent agents for retesting. This not only provides additional information for resolving conflicts but also demonstrates the system's proactive verification and self-correction capabilities, effectively improving the reliability and credibility of the evidence. By constructing a structured evidence network, all evidence and their interrelationships (support, contradiction, verification) are systematically organized and assigned consistency markers. This provides a comprehensive, clear, and highly credible evidentiary foundation for subsequent dynamic graph fusion and traceable reasoning, thereby enhancing the robustness and interpretability of the entire data processing and insight method. It also provides key support for generating accurate comprehensive credibility assessments and interpretable attribution chains.

[0182] In some of the embodiments described above in this application, cross-modal logical consistency verification and numerical compatibility analysis are proposed to identify and label conflicting evidence node pairs. However, in this process, due to the lack of specific spatiotemporal correlation mechanisms and semantic verification based on domain knowledge base, the conflict detection of cross-modal evidence may be incomplete or the labeling may be inaccurate, making it impossible to effectively distinguish different conflict types and intensities, thereby affecting the consistency and reliability of the structured evidence network.

[0183] To address this, this application further proposes a method based on the modal type and content of standardized evidence nodes, performing cross-modal logical consistency verification and numerical compatibility analysis to identify and mark conflicting evidence node pairs that contradict each other. Specifically, the method includes the following steps: Based on the spatial location and acquisition timestamp of standardized evidence nodes, standardized evidence nodes of different modalities are spatiotemporally correlated and paired to form a cross-modal evidence node pair set. This step aims to ensure that evidence of different modalities is correlated in time and space, providing a unified and accurate context for subsequent conflict detection. For example, a time window (e.g., 5 seconds) and a spatial distance threshold (e.g., 10 centimeters) can be set, and different modal evidence nodes acquired within this time window and with spatial distances within the threshold range can be paired. Alternatively, when the system detects a specific event (e.g., a device alarm), all modal evidence nodes within the same physical area (e.g., the same bearing, the same motor housing) of the alarm device component within a specific time period before and after the event are paired. In this way, misjudgments caused by spatiotemporal misalignment can be effectively avoided, ensuring that conflict detection is performed within a unified physical and temporal context.

[0184] For each node pair in the cross-modal evidence node pair set, the evidence content of the standardized evidence nodes in the node pair is parsed, and semantic-level logical consistency verification is performed based on the domain knowledge base of the fault diagnosis domain. This step utilizes domain expert knowledge and historical experience to determine whether the phenomena described by different modal evidence are logically consistent at the semantic level. For example, the domain knowledge base can store a series of logical rules, such as "If the vibration sensor shows abnormal bearing vibration, then thermal imaging should show an increase in bearing temperature." The system uses a rule engine to match and reason about the evidence content to determine whether there are any violations of these preset rules. Another implementation approach is to construct the domain knowledge base as an ontology, defining the semantic relationships between fault modes, symptoms, and sensor readings, and using an ontology inference engine to check whether the semantic descriptions of the evidence node pairs conform to the causal or correlational relationships defined in the ontology.

[0185] For each node pair in the cross-modal evidence node pair set, the numerical features of the standardized evidence nodes in the node pair are extracted, and numerical compatibility analysis is performed based on physical consistency constraints. This step, from a numerical perspective, determines whether the numerical values ​​of different modal evidence conform to physical laws based on physical laws and equipment models. For example, the numerical features of one modality can be used as input to the equipment physical model to predict the theoretical numerical features of another modality, and then compared with the actually acquired numerical features of the other modality to calculate the deviation. Alternatively, physical consistency constraints can define mathematical relationships or allowable numerical ranges between different physical quantities. For example, there may be a fixed proportional relationship between the vibration frequency and rotational speed of a component, or a certain functional relationship between temperature and pressure. The system checks whether the numerical values ​​of the evidence node pairs satisfy these relationships or are within the allowable range.

[0186] Cross-modal evidence node pairs that are identified as contradictory by logical consistency checks or incompatible by numerical compatibility analysis results are marked as conflicting evidence node pairs, and their conflict type and intensity are recorded. This step integrates semantic and numerical analysis results to clearly identify conflicts and quantify their nature and severity, providing a basis for subsequent processing. For example, conflicts can be categorized into different types based on the specific results of logical checks and numerical analysis, such as "semantic contradiction," "numerical exceedance," and "physical model mismatch." Conflict intensity can be quantified using the absolute value of numerical deviation, the percentage of relative deviation, or the severity of logical rule violations. Alternatively, different confidence weights can be assigned to the results of logical checks and numerical analysis, and an overall conflict score can be calculated. When this score exceeds a preset threshold, the node pairs are marked as conflicting evidence node pairs, and the conflict type is determined based on the main factors leading to the high score.

[0187] Through the aforementioned technical solutions, this application can improve the accuracy of conflict detection by introducing a spatiotemporal correlation mechanism to ensure that evidence from different modalities is compared within the correct physical and temporal context. Simultaneously, by combining a domain knowledge base from the fault diagnosis field for semantic-level logical consistency verification, it can effectively identify logical contradictions that are difficult to detect through pure numerical analysis, enhancing the intelligence and accuracy of conflict identification. Furthermore, performing numerical compatibility analysis based on physical consistency constraints ensures the reliability of evidence data at the physical level, improving the credibility of the detection results. Detailed recording of the type and strength of conflicting evidence node pairs provides refined guidance for subsequent targeted cross-validation and conflict resolution, resulting in a structured evidence network with higher consistency and reliability, thus laying a solid foundation for subsequent dynamic graph fusion and traceable reasoning.

[0188] In some of the embodiments described above in this application, conflict detection and cross-validation are proposed to be performed based on real-time acquired evidence during the collection process to form a structured evidence network with consistency markers. However, in this process, when conflicting evidence node pairs are detected, there is a lack of efficient and targeted retesting mechanism to verify the conflict, resulting in insufficient credibility of the evidence network, inability to dynamically resolve contradictions and ensure the spatiotemporal consistency of the evidence.

[0189] To address this, this application further proposes scheduling heterogeneous agents related to the conflict mode, performing retests based on the spatial locations involved in the conflict, generating verification evidence nodes, and updating them to a standardized evidence node set. Specifically, the method includes the following steps: extracting the conflict mode type and the target spatial location involved in the conflict from the conflict evidence node pairs; scheduling an available heterogeneous agent matching the conflict mode type and issuing a directional retest instruction containing the target spatial location to the available heterogeneous agent; receiving the original retest data returned by the available heterogeneous agent, encapsulating the original retest data into verification evidence nodes, each containing evidence content, mode type, target spatial location, and a retest timestamp; and adding the verification evidence nodes to the standardized evidence node set, completing the update of the standardized evidence node set.

[0190] The extraction of conflicting modal types and target spatial locations from conflicting evidence node pairs aims to accurately identify and locate the nature of the evidence conflict and its occurrence area. Conflicting evidence node pairs refer to sets of evidence marked as contradictory in cross-modal logical consistency verification or numerical compatibility analysis. Extracting conflicting modal types, such as vibration, acoustics, and thermal imaging, helps clarify what type of sensor or agent is needed for retesting. Extracting the target spatial locations involved in the conflict, such as the XYZ coordinates of equipment components, sensor numbers, or area identifiers, is used to limit the physical scope of the retesting, ensuring its relevance. This step can directly read the modal type identifiers and spatial location information recorded in the metadata of the conflicting evidence node pairs. For example, each evidence node carries its modal type (e.g., "vibration," "temperature") and acquisition location (e.g., "bearing A, sensor 1, coordinates [x, y, z]") when it is generated. When two nodes conflict, the system can directly extract this information from the metadata of these two nodes. Alternatively, the conflicting modal types and target spatial locations can also be indirectly obtained by querying the original acquisition task records or equipment topology information associated with the conflicting evidence node pairs. For example, if a conflict occurs between vibration data and thermal imaging data of a specific component, the system can query the precise spatial location of the component from the physical model of the equipment based on the component ID associated with these data, and identify the modal types involved.

[0191] Furthermore, a suitable heterogeneous agent matching the conflict mode type is scheduled, and a directional retest instruction containing the target spatial location is issued to the available heterogeneous agents. The aim is to intelligently select and activate the most suitable heterogeneous agent to perform the retest task based on the conflict mode type, and provide it with precise retest instructions. Heterogeneous agents refer to intelligent devices or software agents with different sensing, processing, or execution capabilities, such as vibration sensors, infrared thermal imagers, acoustic sensors, or edge computing units with specific detection algorithms. Matching means that the agent's sensing or detection capabilities conform to the conflict mode type. The directional retest instruction clarifies the executor, target, scope, and method of the retest. This step can be achieved by maintaining a heterogeneous agent capability registry, which records the mode types supported by each agent, its current state (online / offline, busy / idle), and its spatial deployment location. When a retest is needed, the system queries the registry based on the extracted conflict mode type, filters out currently available agents that support that mode type, and selects the optimal one (e.g., the one closest to the target location and with the lowest load) for scheduling. Commands can be issued via wireless communication protocols (such as Wi-Fi, 5G) or industrial buses (such as Ethernet / IP, Profinet). Alternatively, a negotiation mechanism based on agent-based mechanisms can be employed. Upon detecting a conflict, the system issues a retesting task request containing the conflict mode type and target spatial location. Multiple heterogeneous agents assess their ability to respond to this request based on their capabilities and current states, providing feedback on their availability and execution costs to the system. Based on this feedback, the system selects the optimal agent and issues a directional retesting command containing specific acquisition parameters (such as sampling frequency, acquisition duration, and target region of interest) through its API interface or control protocol.

[0192] Based on this, the system receives the original retest data returned by available heterogeneous agents and encapsulates it into verification evidence nodes. These nodes contain evidence content, modality type, target spatial location, and a retest timestamp. The aim is to acquire new data generated after the heterogeneous agents perform the retest and standardize it into a unified evidence node format for subsequent integration into the evidence network. The original retest data is the unprocessed raw sensor data directly collected by the agents. Encapsulating it into verification evidence nodes means structuring the raw data along with its key metadata (such as modality type, acquisition location, and timestamp), making it an identifiable and traceable element in the evidence network. This step can be achieved by the heterogeneous agents uploading the collected raw data (e.g., vibration waveform data, thermal imaging image files) to the data processing platform via a pre-defined data interface (such as MQTT or HTTP POST) after completing the retest. After receiving the data, the platform automatically parses the metadata in the data packet header (such as agent ID and task ID), and combines it with the modality type and target spatial location recorded in the retest instruction, as well as the system timestamp when the data was uploaded, to construct a verification evidence node in JSON or XML format containing this information. Simultaneously, the agent can also perform preliminary processing on the raw data locally (such as format conversion and compression), and package it together with its own generated modality type, precise acquisition spatial coordinates (obtained via GPS or internal positioning system), and local timestamp. Then, this packaged data is sent to the central processing unit through a secure channel. Upon receiving it, the central processing unit either directly uses it as a verification evidence node or performs lightweight verification before encapsulating it.

[0193] Adding verification evidence nodes to the standardized evidence node set updates the standardized evidence node set. This aims to integrate newly generated verification evidence nodes into the existing standardized evidence node set, thereby dynamically updating the evidence network and providing a more comprehensive and accurate evidentiary foundation for subsequent conflict resolution and reasoning. The standardized evidence node set is a collection of all preprocessed and standardized evidence, forming the basis for building a structured evidence network. This step can be achieved by inserting the verification evidence node into a database (such as a time-series database or graph database) storing the standardized evidence node set after its creation. During insertion, the system associates the verification evidence node with existing evidence nodes based on its timestamp and spatial location, and updates relevant indexes to ensure the integrity and queryability of the evidence set. Alternatively, an event-driven mechanism can be used. When a new verification evidence node is generated, the system triggers an update event. Upon receiving this event, the evidence network management module adds the verification evidence node to the in-memory standardized evidence node set data structure and synchronously updates the evidence network topology. For example, it establishes a "verification relationship" edge between the verification evidence node and previously conflicting evidence node pairs, or updates the confidence level of relevant nodes.

[0194] Through the above technical solution, this application provides a targeted retesting mechanism. By precisely scheduling heterogeneous intelligent agents to perform conflict verification and dynamically updating the evidence network, it effectively solves the reliability problem caused by evidence conflicts. Specifically, by extracting the conflict modality type and the target spatial location involved in the conflict from the conflict evidence node pair, the conflict source can be accurately located, avoiding blind retesting and ensuring that the retesting focuses on key contradictions. A usable heterogeneous intelligent agent matching the conflict modality type is scheduled, and a suitable executor is selected based on modality characteristics, improving the accuracy and efficiency of the retesting. Targeted retesting instructions containing the target spatial location are issued to the usable heterogeneous intelligent agent, providing clear operational guidance and tightly binding the retesting process to the conflict location to prevent evidence from becoming detached. The raw retesting data returned by the usable heterogeneous intelligent agent is received, and new evidence is acquired in real time, providing direct input for conflict verification. The raw retesting data is encapsulated into verification evidence nodes, including evidence content, modality type, target spatial location, and retesting timestamp, standardizing the new evidence format to ensure information integrity and traceability. By adding verification evidence nodes to the standardized evidence node set and updating the standardized evidence node set, verification results are dynamically integrated, enhancing the overall consistency of the evidence network and improving diagnostic credibility. This proactive, feedback-based evidence verification mechanism enables the system to move beyond simply identifying conflicts to actively resolving them, thereby enhancing the robustness and reliability of the structured evidence network. This provides a higher-quality evidentiary foundation for subsequent dynamic graph fusion and traceable reasoning, ultimately improving the accuracy and credibility of overall fault diagnosis.

[0195] In some of the solutions mentioned above in this application, dynamic graph fusion and traceable reasoning are proposed to generate comprehensive credibility assessment and explainable attribution chain. However, in the process of implementation, the fusion of evidence network and knowledge path may become static, resulting in unclear reasoning path, single credibility assessment, and inability to effectively handle evidence conflict and ensure the traceability of attribution chain.

[0196] To address this, this application proposes a method for dynamic graph fusion and traceable reasoning of the structured evidence network and the associated knowledge subgraph path to generate a comprehensive credibility assessment and an explainable attribution chain. (See [link to relevant documentation]). Figure 7 This includes the following steps.

[0197] 701. The standardized evidence nodes in the structured evidence network are used as temporary nodes and dynamically inserted into the path of the associated knowledge subgraph to construct a fusion reasoning graph.

[0198] The structured evidence network is a graph structure composed of multiple standardized evidence nodes and their interrelationships (such as support, contradiction, and verification). It reflects the internal consistency and conflict of multimodal evidence acquired in real time. Each standardized evidence node represents a piece of evidence that has undergone modality resolution and spatiotemporal synchronization alignment, containing information such as evidence content, modality type, spatial location, and acquisition timestamp. The associated knowledge subgraph path is a knowledge path retrieved from the domain knowledge graph and related to the initial fault hypothesis, containing prior knowledge such as historical experience rules, fault correlations, and equipment topology. Inserting standardized evidence nodes as "temporary nodes" means that these evidence nodes are only introduced in this reasoning session. They are not permanently modified or stored in the original associated knowledge subgraph path, but rather provide immediate and real-time evidence support for the current fault diagnosis task, ensuring the stability of the knowledge graph and the timeliness of the evidence. This "dynamic insertion" can be achieved in several ways: One approach is based on semantic matching algorithms, which matches the content of standardized evidence nodes (e.g., equipment components, fault phenomena, physical quantities) with entity nodes or event nodes in the associated knowledge subgraph path, and establishes new edges at the matching points to connect the evidence nodes to the knowledge graph. Another approach is to align the evidence nodes with the spatiotemporal range of the equipment components or events described in the associated knowledge subgraph path based on their spatiotemporal information, thereby determining the insertion position and establishing edges representing relationships such as "observed" or "supported." In this way, real-time evidence and prior knowledge are combined into a unified graph model, namely the fused reasoning graph, which contains fault hypotheses, physical model information, historical knowledge, real-time evidence, and their interrelationships, providing a comprehensive information view for subsequent graph structure reasoning.

[0199] 702. Perform graph structure reasoning on the fused reasoning graph, identify the optimal reasoning path connecting key evidence nodes and candidate fault nodes, and convert the optimal reasoning path into the explainable attribution chain.

[0200] In this context, graph-based reasoning refers to using graph algorithms and reasoning mechanisms on a fused reasoning graph to discover hidden patterns, relationships, or paths. For example, path-search-based algorithms, such as A* search, Dijkstra's algorithm, or depth-first / breadth-first search, combined with heuristic functions, can be used to find paths from evidence to faults. Alternatively, knowledge graph reasoning techniques, such as rule-based reasoning, embedded reasoning, or graph neural networks (GNNs), can be used to identify the logical connections between evidence and faults. The key evidence node refers to a standardized evidence node in the structured evidence network that has high confidence, strong supporting role, or is directly related to the fault hypothesis; these nodes are the core support for diagnostic reasoning. The candidate fault node represents various fault hypotheses in the fault hypothesis space to be verified and is the target of the reasoning. The optimal reasoning path is the path with the highest score among all possible evidence-to-fault paths, based on preset evaluation criteria (such as path length, node confidence, edge weight, semantic relevance, etc.). Path scoring functions can comprehensively consider the confidence of all evidence nodes on the path, the weight of edges between knowledge nodes and evidence nodes, and the topological complexity of the path (such as path length and number of branches) to select the path with the highest score. Alternatively, a path evaluator can be trained using a machine learning model to predict the "optimality" of a path based on its characteristics. This optimal reasoning path is then transformed into an interpretable attribution chain, presenting the optimal reasoning path in a human-understandable form (usually natural language text), clearly demonstrating the causal logical chain from observed evidence to the conclusion of failure. This can be achieved by using predefined templates and Natural Language Generation (NLG) techniques to map the node and edge relationships in the optimal reasoning path into descriptive statements, combining them into a coherent explanatory text. Alternatively, semantic parsing and Knowledge Graph Question Answering (KGQA) techniques can be used to transform the reasoning path into a structured semantic representation, and then a text generation model can be used to generate a detailed attribution explanation.

[0201] 703. Based on the node relationships in the fusion reasoning graph, the consistency markers in the structured evidence network, and the physical consistency constraints in the reasoning skeleton, the comprehensive credibility assessment is determined. The comprehensive credibility assessment integrates physical consistency confidence, multimodal consistency confidence, and knowledge graph support.

[0202] The node association refers to the semantic relationship and strength between nodes in the fusion reasoning graph, as embodied by edges. Its weight or type reflects the strength or nature of the association. The consistency label is a marker on each standardized evidence node or node pair in the structured evidence network, indicating whether there is conflict or mutual support between pieces of evidence. For example, it can be in states such as "consistent," "conflicting," or "to be verified," and may include a conflict strength measure. The physical consistency constraint originates from the equipment's physical model, defining the physical laws, engineering principles, or numerical ranges that physical quantities should satisfy under specific failure modes. It is used to assess whether the evidence or failure hypothesis conforms to the objective laws of the physical world. The comprehensive credibility assessment is a quantitative indicator used to measure the overall reliability of the current failure diagnosis conclusion. It integrates confidence levels from three aspects: physical, multimodal evidence, and knowledge graph. Physical consistency confidence measures the degree to which evidence or failure hypothesis conforms to physical consistency constraints. Multimodal consistency confidence measures the degree of consistency and mutual support between different modalities of evidence in the structured evidence network. Knowledge graph support measures the strength of knowledge support obtained by the failure hypothesis or reasoning path in the associated knowledge subgraph path. These three confidence metrics can be combined in several ways. For example, a weighted average method can be used, assigning different weights to each confidence metric and then summing them to obtain the overall confidence score. Alternatively, Bayesian networks, fuzzy logic, or machine learning models can be used, taking these three confidence scores as input features and outputting an overall confidence score assessment.

[0203] Through the above technical solution, this application effectively solves the static problem in the fusion of evidence networks and knowledge paths. By dynamically inserting standardized evidence nodes into the paths of associated knowledge subgraphs, a fusion reasoning graph is constructed, achieving a close integration of real-time evidence and prior knowledge. This avoids the problems of evidence "floating" and the "physical decoupling" and "logical defocusing" of knowledge retrieval, ensuring that the reasoning process can respond to changes in evidence in real time. Graph structure reasoning is performed on the fusion reasoning graph, identifying and converting it into an interpretable attribution chain. This makes the causal logic chain from evidence to fault clear and transparent, solving the problems of unauditable reasoning processes and untraceable attribution chains in traditional "black box" models. Simultaneously, based on the node relationships in the fusion reasoning graph, the consistency markers within the structured evidence network, and the physical consistency constraints in the reasoning skeleton, a comprehensive credibility assessment is determined. This assessment integrates physical consistency confidence, multimodal consistency confidence, and knowledge graph support, thus providing a multi-dimensional and more comprehensive credibility quantification. This avoids the singleness of assessment and can proactively identify and handle evidence conflicts, improving the reliability and credibility of fault diagnosis conclusions. This dynamic fusion, traceable reasoning, and comprehensive evaluation mechanism enables the system to provide more accurate and convincing diagnostic results when facing complex and compound faults, providing a solid foundation for subsequent closed-loop operation and maintenance decisions.

[0204] In some of the embodiments described above in this application, graph structure reasoning is proposed to perform graph structure reasoning on a fused reasoning graph to identify the optimal reasoning path and convert it into an interpretable attribution chain to generate a transparent causal reasoning chain. However, in its implementation, path search may be inefficient and inaccurate, path scoring lacks a unified mechanism that comprehensively considers node confidence, edge weight and topological complexity, and the interpretation conversion may be unnatural or unreliable, making the reasoning process difficult to audit and the conclusions unreliable.

[0205] To address this, this application further proposes a method for performing graph structure reasoning on a fused reasoning graph to identify the optimal reasoning path connecting key evidence nodes and candidate fault nodes, and converting the optimal reasoning path into an interpretable attribution chain. This method includes: using key evidence nodes as the source node set and candidate fault nodes as the sink node set, performing a graph traversal search on the fused reasoning graph to generate a set of candidate reasoning paths connecting the source and sink node sets. Based on a preset path scoring function, calculating the path score for each candidate reasoning path in the set, where the path score integrates the confidence of nodes, the association weight of edges, and the topological complexity of the candidate reasoning path. Selecting the candidate reasoning path with the highest path score from the set as the optimal reasoning path. Analyzing the node attributes and edge relationships in the optimal reasoning path to generate a causal logical chain representing the transition from evidence to fault, and converting the causal logical chain into an interpretable attribution chain in natural language text form.

[0206] Specifically, using key evidence nodes as the source node set and candidate fault nodes as the sink node set, a graph traversal search is performed in the fused inference graph to generate a set of candidate inference paths connecting the source and sink node sets. This step aims to limit the scope and objective of the graph traversal search, thereby efficiently discovering all possible causal paths from observed evidence to potential faults. By explicitly defining the starting point (key evidence nodes) and the ending point (candidate fault nodes), blind or inefficient searches across the entire fused inference graph can be avoided, improving the efficiency and relevance of path discovery. In one implementation, breadth-first search (BFS) or depth-first search (DFS) algorithms can be used. For example, starting from each key evidence node, the graph is traversed along its edges until any candidate fault node is reached. During the traversal, all traversed nodes and edges are recorded, forming a complete path. To avoid loops and repeated paths, a list of visited nodes or a set of paths can be maintained. In another implementation, the A* search algorithm can be used, combined with a heuristic function to guide the search direction, making it more inclined to discover "promising" paths. Heuristic functions can be based on information such as node type, edge weight, or estimated path length. This approach can more effectively prune and accelerate the discovery of the optimal path when the graph structure is complex and the number of paths is large.

[0207] Based on this, a path score is calculated for each candidate inference path in the candidate inference path set using a preset path scoring function. The path score comprehensively considers the confidence of nodes, the association weights of edges, and the topological complexity of the candidate inference path. The core of this step is to quantitatively evaluate each discovered candidate inference path to determine its rationality and credibility as the "optimal" path. By comprehensively considering the confidence of nodes, the association weights of edges, and the topological complexity of the path, the quality of a path can be comprehensively measured, avoiding biases caused by a single indicator and ensuring that the selected optimal path performs excellently in terms of evidence support, knowledge association, and structural simplicity. In one implementation, the path scoring function can be designed as a weighted sum of its components. For example, path score = w1 × (sum of node confidence) + w2 × (sum of edge association weights) - w3 × (topological complexity). Node confidence can be derived from consistency markers of evidence or reliability assessments of nodes in the knowledge graph. Edge association weights can be derived from the strength of relationships in the knowledge graph or the causal strength in the physical model. Topological complexity can be simply represented as the number of hops or nodes in a path. Weights w1, w2, and w3 can be determined through training based on domain expert experience or machine learning methods. Alternatively, path scoring can be implemented using multi-objective optimization methods, treating node confidence, edge association weights, and topological complexity as different optimization objectives. For example, paths that meet certain confidence and weight thresholds can be selected first, and then the path with the lowest topological complexity can be chosen from these paths.

[0208] The optimal inference path is selected from the candidate inference path set, based on the highest path score. This step is crucial for decision-making based on quantitative evaluation results, aiming to clearly identify the path that best explains the current fault phenomenon and has the highest credibility from numerous possible causal paths. By directly selecting the highest-scoring path, the uniqueness and optimality of the inference result are ensured, providing a solid foundation for subsequent attribution chain generation. In one implementation, after calculating the path scores of all candidate inference paths, the system can directly traverse the entire set to find the path with the highest score. In another implementation, all candidate inference paths can be sorted in descending order of their path scores, and then the first path in the sorted list can be selected.

[0209] This process involves analyzing the node attributes and edge relationships in the optimal reasoning path to generate a causal logic chain representing the path from evidence to fault, and then converting this causal logic chain into an interpretable attribution chain in natural language text form. This step aims to transform the abstract graph-structured reasoning results into human-understandable and auditable fault diagnosis explanations. By analyzing the semantic information of nodes (representing evidence, intermediate events, fault modes, etc.) and edges (representing causal relationships, correlations, etc.) on the optimal path, a clear causal logic chain is constructed, and further converted into natural language. This significantly improves the interpretability and credibility of the diagnostic conclusions, facilitating understanding and decision-making by engineers. In one implementation, a series of natural language templates can be predefined, each template corresponding to a node type or edge relationship. For example, when encountering an edge "Evidence A leads to event B," the template "Due to [the description of evidence A], [the description of event B] was caused" can be filled in. By mapping the node attributes and edge relationships in the optimal path to these templates and concatenating them sequentially, a complete natural language attribution chain can be generated. In another implementation, a Natural Language Generation (NLG) model can be used, taking the structured representation of the optimal reasoning path as input, and training the model to directly generate fluent and accurate natural language descriptions.

[0210] Through the above technical solution, this application defines key evidence nodes as the source node set and candidate fault nodes as the sink node set, and performs graph traversal search in the fused inference graph, effectively limiting the search scope and avoiding blind traversal, thereby improving the efficiency and accuracy of path search and solving the problem of low path search efficiency. Simultaneously, based on a preset path scoring function, it comprehensively considers the confidence of nodes in candidate inference paths, the association weight of edges, and the topological complexity of paths, achieving multi-dimensional and comprehensive path quality assessment, overcoming the limitations of the single scoring mechanism of traditional methods, and ensuring the accuracy and reliability of optimal path selection. By selecting the candidate inference path with the highest path score as the optimal inference path, the decision-making process is simplified and inference efficiency is improved. Furthermore, by parsing the node attributes and edge relationships in the optimal inference path, a causal logical chain representing the path from evidence to fault is generated and converted into natural language text, making the interpretation of diagnostic conclusions more natural, intuitive, and easy to understand, solving the problem of unnatural or unreliable interpretation conversion. This not only enhances the auditability of the inference process but also greatly improves the credibility of diagnostic conclusions, providing engineers with clear and transparent fault attribution basis. This approach works synergistically with the aforementioned technical solution that dynamically fuses structured evidence networks with associated knowledge subgraph paths, enabling the fused reasoning graph to provide rich and accurate contextual information. This lays a solid foundation for efficient path search and accurate scoring in this approach, ensuring that the generated interpretable attribution chain has both physical consistency and multimodal evidence support, as well as deep association with the knowledge graph.

[0211] In some of the embodiments described above in this application, a comprehensive credibility assessment based on fusion reasoning graph generation is proposed to improve the credibility of fault diagnosis. However, in its implementation, due to the lack of quantitative calculation and weighted fusion of physical feature matching degree, cross-modal evidence synergy status and knowledge node support, the credibility assessment is singular and unreliable, and cannot fully reflect the physical consistency, multimodal synergy and knowledge graph support of the evidence, thereby reducing the accuracy and interpretability of the diagnostic conclusion.

[0212] To address this, this application further proposes a method for determining the comprehensive credibility assessment. This method is based on the node relationships in the fusion inference graph, the consistency markers within the structured evidence network, and the physical consistency constraints in the inference skeleton. Specifically, the method includes the following steps: Based on the physical consistency constraints defined in the inference skeleton, assess the degree of physical feature matching of standardized evidence nodes in the structured evidence network, and calculate the physical consistency confidence. Based on the consistency markers within the structured evidence network, quantify the collaborative support state and conflict degree between cross-modal evidence nodes, and calculate the multimodal consistency confidence. Based on the edge weights and path support between standardized evidence nodes and knowledge nodes in the fusion inference graph, calculate the knowledge graph support. According to a preset weight allocation strategy, weighted fusion of the physical consistency confidence, the multimodal consistency confidence, and the knowledge graph support generates the comprehensive credibility assessment.

[0213] The physical consistency constraint is structured metadata derived from the equipment physical model deduction logic. It defines the physical laws, causal relationships, or state boundary conditions that the various physical quantities of the equipment should satisfy under specific fault assumptions, providing a benchmark for the physical rationality of the evidence. This structured evidence network is a dynamically constructed graph structure, where nodes represent standardized evidence data (standardized evidence nodes), and edges represent the supporting, contradictory, or verifying relationships between these evidence nodes. It aims to integrate original evidence from different modalities, times, and spatial locations, and to mark it for consistency. The standardized evidence node is a unified representation of the original multimodal evidence data after preprocessing, encapsulating the core information of the evidence, including its content, modality type, spatial location at the time of collection, and precise timestamp. The physical feature matching degree refers to the degree of agreement between the actual observed physical features contained in the standardized evidence node and the physical consistency constraints defined in the inference skeleton, quantifying the consistency between actual evidence and the predictions of the theoretical physical model. The physical consistency confidence is calculated based on the physical feature matching degree, reflecting the reliability of the current evidence at the physical level in conforming to the equipment model and physical laws. The consistency marker is metadata used in the structured evidence network to identify the cooperative support or contradictory states among evidence nodes. It is generated through conflict detection and cross-validation mechanisms, explicitly indicating which evidence pairs have logical conflicts or numerical incompatibilities. The cross-modal evidence node refers to standardized evidence nodes from different sensing modalities that describe the same device component, the same time period, or the same physical phenomenon; they are formed through spatiotemporal pairing. The cooperative support state refers to the degree to which different modal evidence mutually corroborates, complements, or enhances each other. The conflict degree refers to the degree of contradiction or inconsistency among them; these states and degrees are obtained through the parsing and quantification of the consistency marker. The multimodal consistency confidence score is calculated based on the cooperative support state and conflict degree among cross-modal evidence nodes, reflecting the reliability of mutual corroboration and logical consistency among different modal evidence. The fusion reasoning graph is constructed by dynamically inserting standardized evidence nodes from the structured evidence network onto the path of the associated knowledge subgraph, integrating real-time observational evidence from the physical world with prior knowledge, historical experience, and logical relationships in the domain knowledge graph. This knowledge node is a fundamental building block in the domain knowledge graph, representing entities such as equipment components, failure modes, phenomena, causes, and maintenance measures, as well as the various relationships between them. The associated edge weight refers to the importance or reliability measure of the edges connecting knowledge nodes or evidence nodes in the fusion reasoning graph. Path support is a comprehensive reflection of all edge weights and node confidence on a reasoning path from an evidence node to a candidate failure node. This knowledge graph support is calculated based on the associated edge weights and path support between standardized evidence nodes and knowledge nodes in the fusion reasoning graph, reflecting the degree to which the diagnostic conclusion or failure hypothesis is supported by the domain knowledge graph.The pre-defined weighting strategy refers to the relative importance coefficients assigned to each confidence component when weighted and fused with physical consistency confidence, multimodal consistency confidence, and knowledge graph support. These weights can be adjusted based on the actual application scenario, equipment type, fault diagnosis stage, or expert experience. This weighted fusion is a common method for combining multiple indicators or scores with different importance into a single comprehensive score. The comprehensive credibility assessment is a quantitative indicator output by this method that reflects the overall reliability of the fault diagnosis conclusion or fault hypothesis.

[0214] Through the above technical solutions, this application introduces a multi-source confidence calculation and weighted fusion mechanism, solving the problem of unreliable single confidence assessment, thereby improving the comprehensiveness and reliability of fault diagnosis. Specifically, it evaluates the degree of physical feature matching of standardized evidence nodes in the structured evidence network based on the physical consistency constraints defined in the inference skeleton, and calculates the physical consistency confidence. This utilizes the inherent logic of physical constraints to quantify the matching degree between evidence and device models, effectively solving the "physical decoupling" problem and ensuring that the assessment is based on real physical reality. Simultaneously, it quantifies the collaborative support state and conflict degree between cross-modal evidence nodes based on consistency tags within the structured evidence network, and calculates the multimodal consistency confidence. This captures the collaborative or conflicting relationships between evidence through the dynamic feedback mechanism of consistency tags, solving the inconsistency problem in multimodal fusion and enhancing evidence synergy. Furthermore, it calculates the knowledge graph support based on the association edge weights and path support between standardized evidence nodes and knowledge nodes in the fusion inference graph. This incorporates historical experience knowledge by leveraging the association weights and path support of knowledge nodes, solving the problem of knowledge update lag and improving the knowledge support of the assessment. Based on a pre-defined weighting strategy, physical consistency confidence, multimodal consistency confidence, and knowledge graph support are weighted and fused to generate a comprehensive credibility assessment. This dynamic integration of multi-dimensional confidence through the weighting strategy solves the problem of singular credibility and achieves a comprehensive and reliable assessment output. Overall, this multi-dimensional, weighted fusion assessment method ensures that diagnostic conclusions are supported not only by real-time evidence but also by physical laws and historical experience, improving the accuracy, credibility, and interpretability of the diagnostic conclusions and providing a solid foundation for the subsequent generation of closed-loop optimization instructions.

[0215] In some of the solutions mentioned above in this application, a data processing and insight method based on a multimodal intelligent assistant is proposed. However, in this process, when the overall credibility assessment of the diagnostic conclusion is lower than the preset threshold or there are unresolved conflicts in the structured evidence network, there is a lack of a mechanism to automatically trigger deep reflection and generate closed-loop optimization instructions. This results in the system being unable to actively supplement verification or update the knowledge model, leading to insufficient credibility of the diagnostic conclusion and a lack of self-evolution capability of the system.

[0216] In response, this application further proposes a method for generating closed-loop optimization instructions based on comprehensive credibility assessment and consistency labeling-triggered metacognitive reflection. (See [link to relevant documentation]). Figure 8 The method includes the following steps: 801. Compare the comprehensive credibility assessment with the preset credibility threshold, and analyze the consistency markers in the structured evidence network. Based on the comparison and analysis results, determine whether metacognitive reflection is triggered.

[0217] This step aims to intelligently determine whether a deeper self-reflection process needs to be initiated by dually assessing the confidence level of the current diagnostic conclusion and the internal consistency of the collected evidence. The comprehensive credibility assessment is a quantitative measure of the overall reliability of the diagnostic result. Comparing it to a preset credibility threshold can preliminarily determine whether the current diagnostic conclusion has reached the expected confidence level. For example, this comparison can be a simple numerical judgment, where a value below the threshold indicates potential uncertainty in the diagnostic conclusion. Alternatively, an interval judgment can be used, triggering reflection when the value falls into a certain uncertain interval. Various consistency markers contained in the structured evidence network reflect the support, contradiction, or verification relationships between different modalities of evidence, and between evidence and prior knowledge. Analyzing these markers aims to identify unresolved conflicts, contradictions, or insufficient evidence in the chain of evidence. For example, it can analyze whether there are evidence node pairs marked as "conflicting" in the network, or whether there is isolated evidence lacking sufficient support. Another approach is to assess the severity and scope of conflicts in the evidence network. The judgment result is based on a comprehensive consideration of the above comparisons and analyses. If the overall credibility of the diagnostic conclusion is insufficient, or if there are significant unresolved conflicts in the evidence network, the system will determine that metacognitive reflection needs to be triggered. This ensures that the reflection process is initiated only when truly needed, avoiding unnecessary resource consumption, while guaranteeing a timely response to potential problems.

[0218] 802. When metacognitive reflection is triggered, obtain the reasoning skeleton, structured evidence network, related knowledge subgraph path and interpretable attribution chain related to the current diagnostic task as reflection context data.

[0219] Once the system determines that metacognitive reflection is necessary, it collects all key information about the current diagnostic task to construct a comprehensive reflective context. This ensures that subsequent reflection is based on a complete and accurate diagnostic history and status. The reasoning skeleton defines the initial hypothesis space, evidence requirements, and constraints for this diagnosis. The structured evidence network contains all collected, processed, and verified multimodal evidence and their relationships. The associated knowledge subgraph provides domain knowledge and historical experience relevant to the current failure situation. The interpretable attribution chain demonstrates the logical derivation path from evidence to failure hypotheses. These data collectively constitute a complete description of the current diagnostic task. Aggregating these key data sets forms a unified reflective context data set. This allows the reflective module to comprehensively examine all aspects of the diagnostic process, including the formation of initial hypotheses, evidence collection and integration, knowledge application, and conclusion drawing, providing the necessary foundation for in-depth analysis and problem localization.

[0220] 803. Based on reflective context data, perform an audit of the completeness of the diagnostic assumption coverage and an assessment of the logical reliability of the core chain of evidence, and generate a reflective assessment report.

[0221] After acquiring comprehensive reflective context data, the system will conduct an in-depth self-examination of its diagnostic process, focusing on assessing the comprehensiveness of diagnostic hypotheses and the rigor of the reasoning process, and summarizing the review results into a reflective evaluation report. This audit aims to check whether the system has adequately considered all possible failure modes and potential causes in the current diagnostic task. For example, the system can compare the space of unverified failure hypotheses defined in the inference framework with historical failure cases or a broader knowledge base contained in the associated knowledge subgraph paths to identify potential failure modes not covered by the current diagnostic process but relevant to the current symptoms. This helps to discover blind spots or knowledge gaps in the diagnostic process. The evaluation focuses on the derivation path from evidence to failure represented by the interpretable attribution chain. The system will examine whether the logical relationships between evidence nodes supporting the attribution chain in the structured evidence network are rigorous and whether there are contradictions, and verify whether the entire derivation process conforms to the pre-set physical consistency constraints in the inference framework. For example, it can check whether the supporting relationships between evidence nodes are strong enough, or whether there are logical jumps or inconsistencies. The audit and evaluation results will be structurally integrated to form a reflective evaluation report. This report will clearly identify potential gaps in the diagnostic hypotheses and logical conflicts in the core chain of evidence, providing corresponding confidence scores or severity ratings for each issue. This report serves as a crucial basis for generating subsequent closed-loop optimization instructions.

[0222] 804. Based on the reflection and evaluation report, generate closed-loop optimization instructions. These instructions include immediate verification instructions for scheduling supplementary evidence collection, or long-term evolution instructions for updating the domain knowledge graph and device physical model.

[0223] Based on the issues and suggestions identified in the reflection and evaluation report, the system will automatically generate specific, executable optimization instructions to achieve immediate correction of the diagnostic process or long-term improvement of the system's knowledge model. Closed-loop optimization instructions are concrete action plans for the system's self-correction and evolution. These instructions are tailored to the issues identified in the reflection and evaluation report, ensuring the targeted nature and effectiveness of the optimization measures. If the reflection and evaluation report indicates insufficient current evidence, unresolved conflicts, or a lack of sufficient support for a key hypothesis, the system will generate an immediate verification instruction. This instruction is used to schedule heterogeneous agents to collect supplementary evidence to obtain more, more accurate, or different modal evidence to verify or eliminate specific fault hypotheses. For example, the instruction might require additional thermal imaging scans or vibration data collection for a specific area. If the reflection and evaluation report reveals structural defects, knowledge gaps, or model inaccuracies in the domain knowledge graph or equipment physical model (e.g., discovering new fault modes that have not been included, or deviations in the physical model's simulation of certain operating conditions), the system will generate long-term evolution instructions. These instructions are designed to guide human experts or automated learning modules to update, revise, or expand the domain knowledge graph and device physical model, thereby improving the system's future diagnostic capabilities and adaptability.

[0224] Through the aforementioned technical solution, this application introduces a metacognitive reflection mechanism, addressing the problem that the system lacks the ability to automatically trigger deep reflection and generate closed-loop optimization instructions when the overall credibility assessment of the diagnostic conclusion falls below a preset threshold or unresolved conflicts exist in the structured evidence network. Specifically, by comparing the overall credibility assessment with a preset credibility threshold and parsing consistency markers in the structured evidence network, the system can intelligently determine when to initiate deep reflection, avoiding unnecessary resource consumption while ensuring timely responses to potential problems. When metacognitive reflection is triggered, the system acquires the reasoning framework, structured evidence network, associated knowledge subgraph paths, and interpretable attribution chains related to the current diagnostic task as reflection context data. This ensures that subsequent audits and evaluations are based on a comprehensive and accurate diagnostic history and status, thereby improving the effectiveness of reflection. Based on this, the system performs an audit of the completeness of the diagnostic hypothesis coverage and an evaluation of the logical reliability of the core evidence chain, generating a reflection evaluation report. This enables the system to identify blind spots, knowledge gaps, or illogical inferences in the diagnostic process, enhancing the reliability and credibility of the diagnostic results. Based on the reflection and evaluation report, the system generates closed-loop optimization instructions, including immediate verification instructions for scheduling supplementary evidence collection and long-term evolutionary instructions for updating the domain knowledge graph and device physical model. This closed-loop feedback mechanism enables the system to proactively correct the uncertainty of the current diagnosis and achieve self-evolution through continuous learning and knowledge updates. This fundamentally solves the problems of insufficient credibility of diagnostic conclusions and lack of system self-evolution capabilities, improving the robustness and adaptability of the intelligent diagnostic system.

[0225] In some of the embodiments described above in this application, a metacognitive reflection based on comprehensive credibility assessment and consistency markers is proposed to determine when to conduct reflection. However, in this process, the triggering conditions may lack precise quantitative standards and conflict depth analysis, which may lead to reflection being triggered unnecessarily or not being triggered in a timely manner when key evidence conflicts, affecting system efficiency and diagnostic reliability.

[0226] To address this, this application further proposes comparing the comprehensive credibility assessment with a preset credibility threshold and analyzing consistency markers in the structured evidence network. Based on the comparison and analysis results, it determines whether metacognitive reflection is triggered. Specifically, this determination process includes: comparing the value of the comprehensive credibility assessment with the preset credibility threshold; analyzing consistency markers in the structured evidence network to identify whether unresolved conflicts exist in the structured evidence network and determining the conflict intensity of the unresolved conflicts; and determining whether metacognitive reflection is triggered if the value of the comprehensive credibility assessment is lower than the preset credibility threshold, or if unresolved conflicts with a conflict intensity higher than the preset conflict intensity threshold are identified.

[0227] The overall credibility assessment value is a quantitative representation of the system's overall credibility of the current diagnostic conclusion. It integrates multiple dimensions such as physical consistency confidence, multimodal consistency confidence, and knowledge graph support, and is typically between 0 and 1, with higher values ​​indicating higher credibility. The preset credibility threshold is a pre-defined benchmark value used to judge whether the diagnostic conclusion is sufficiently credible. Its setting can be based on statistical analysis of historical diagnostic data, such as analyzing the average credibility of past successful diagnostic cases, or determined according to the accuracy requirements of specific application scenarios (e.g., high-risk equipment fault diagnosis may require a higher threshold), for example, it can be set to 0.7 or 0.85. Comparing the overall credibility assessment value with the preset credibility threshold aims to quickly determine whether the current diagnostic result has reached the expected credibility level by directly comparing the actually calculated overall credibility assessment value with the preset threshold.

[0228] Simultaneously, the consistency markers in the structured evidence network are analyzed to identify unresolved conflicts and determine their intensity. The structured evidence network is formed during the collaborative evidence gathering phase, where nodes represent standardized evidence and edges represent supporting, contradictory, or verifying relationships between pieces of evidence. Consistency markers are the results of evaluating evidence consistency through conflict detection and cross-validation mechanisms during evidence gathering; they indicate whether logical contradictions or numerical incompatibilities exist between evidence nodes. Unresolved conflicts refer to evidentiary contradictions that could not be resolved through cross-validation or fusion algorithms during the evidence gathering and preliminary verification phases. Identifying these conflicts is crucial for ensuring diagnostic reliability. The identification process may include traversing the structured evidence network, finding all edges marked as contradictory relationships, and checking whether these contradictions have been reconciled by subsequent verifying evidence nodes. Conflict intensity is a quantitative indicator of the severity of unresolved conflicts. For example, conflict intensity can be calculated based on factors such as the numerical differences between conflicting evidence node pairs, the severity of logical contradictions (e.g., direct negation versus indirect inconsistency), the number of evidence modalities involved, and the criticality of the conflicting evidence in the reasoning path. For example, it can be determined by calculating the Euclidean distance between conflicting evidence node pairs, the inverse of semantic similarity, or a weighted score based on expert rules. The preset conflict intensity threshold is a pre-defined benchmark value used to determine whether unresolved conflicts are severe enough to trigger metacognitive reflection. Its setting can be based on the system's conflict tolerance requirements; for example, minor conflicts that do not affect the core diagnosis can be tolerated, while conflicts involving critical evidence or high-risk failure modes require reflection. For instance, it could be set to 0.6 or 0.75.

[0229] If the overall credibility assessment value is lower than the preset credibility threshold, or if an unresolved conflict with a conflict intensity higher than the preset conflict intensity threshold is identified, then metacognitive reflection is triggered. This judgment logic uses an "OR" relationship, meaning that as long as either condition is met, the current diagnostic process is considered to have a significant problem, and the metacognitive reflection mechanism needs to be activated.

[0230] Through the aforementioned technical solution, this application achieves precise control over the timing of metacognitive reflection triggering, effectively avoiding unnecessary reflection operations while ensuring timely intervention at critical moments. Specifically, by comparing the comprehensive credibility assessment value with a preset credibility threshold, the system can judge the overall reliability of the current diagnostic conclusion based on objective quantitative standards, avoiding the abuse of reflection due to subjective judgment or vague standards, thereby improving diagnostic efficiency. Furthermore, by deeply analyzing the consistency markers in the structured evidence network and identifying and quantifying the intensity of unresolved conflicts, this application can directly conduct in-depth analysis of contradictions at the evidentiary level. When unresolved conflicts with a conflict intensity exceeding the preset threshold are detected, the system can trigger reflection in a timely manner even if the overall credibility is acceptable. This allows the system to keenly capture potential key evidentiary issues that may lead to misjudgment. This dual judgment mechanism, combining overall credibility with the severity of local evidentiary conflicts, makes the triggering of metacognitive reflection more intelligent and robust, ensuring that the system only initiates resource-intensive reflection processes when truly needed, improving the accuracy, reliability, and resource utilization efficiency of the diagnosis.

[0231] In some of the solutions described above in this application, a metacognitive reflection based on comprehensive credibility assessment and consistency marking is proposed to generate closed-loop optimization instructions, which are used to proactively plan supplementary verification or update the knowledge base when there is evidence conflict or high uncertainty in the conclusion. However, in this process, metacognitive reflection may lack a systematic mechanism to audit the coverage completeness of diagnostic hypotheses (i.e., whether all potential failure modes are taken into consideration) and assess the logical reliability of the core evidence chain (i.e., whether the evidence chain is logically consistent and conforms to physical constraints), resulting in incomplete or inaccurate reflection and evaluation, which in turn affects the accuracy and executability of the closed-loop optimization instructions.

[0232] To address this, this application further proposes a method based on reflective context data to audit the completeness of diagnostic hypothesis coverage and assess the logical reliability of the core evidence chain, generating a reflective assessment report. This process includes the following steps: comparing the unverified fault hypothesis space in the inference framework with historical fault case nodes in the associated knowledge subgraph paths to identify potential fault modes not covered by the associated knowledge subgraph paths, generating a coverage completeness audit result. Based on consistency markers in the structured evidence network, verifying the logical support relationships between evidence nodes upon which the interpretable attribution chain depends. Verifying whether the overall derivation process of the interpretable attribution chain conforms to the physical consistency constraints defined in the inference framework, generating a logical reliability assessment result. Integrating the coverage completeness audit result and the logical reliability assessment result, generating a reflective assessment report, which includes identifiers of missing coverage items, identifiers of logical conflicts, and confidence scores for each item.

[0233] Specifically, the step of comparing the fault hypothesis space to be verified in the reasoning skeleton with historical fault case nodes in the associated knowledge subgraph path to identify potential fault modes not covered by the associated knowledge subgraph path and generate coverage integrity audit results aims to assess whether the fault hypotheses considered in the current diagnostic process are comprehensive and whether there are any omitted potential fault modes. One implementation is to use a comparison method based on semantic similarity. For example, for each fault hypothesis in the fault hypothesis space to be verified, its key features (such as fault type, component, phenomenon) are extracted and semantically matched with the fault description in the historical fault case node. If the semantics of a historical fault case node does not match any fault hypothesis in the fault hypothesis space to be verified or the similarity is lower than a preset threshold, it is identified as an uncovered potential fault mode. Another implementation is to use a comparison method based on graph structure topology. The fault modes in the fault hypothesis space to be verified are mapped to nodes in the associated knowledge subgraph path, and then the historical fault case nodes in the associated knowledge subgraph path that are not mapped but have high importance or relevance are analyzed. These unmapped nodes are considered as potential uncovered failure modes.

[0234] In the step of verifying the logical support relationships between evidence nodes upon which an interpretable attribution chain depends, based on consistency markers in a structured evidence network, the aim is to evaluate the internal logical consistency and reliability of the interpretable attribution chain. By utilizing existing consistency markers in the structured evidence network (e.g., indicating whether the relationship between evidence nodes is supportive, contradictory, or confirmatory), it is possible to check for logical conflicts or insufficient support among the evidence nodes in the interpretable attribution chain. One implementation is to traverse each evidence node and its adjacent nodes in the interpretable attribution chain, querying the consistency markers between these nodes in the structured evidence network. If a contradictory relationship marker is found between an evidence node and its next supported node, or if a clear support relationship marker is lacking, the logical support relationship is considered problematic. Another implementation is to construct a set of logical reasoning rules. The consistency markers in the structured evidence network are converted into formalized logical rules, and then the interpretable attribution chain is input as a proposition to be verified into a logical reasoning engine. If the reasoning engine finds that a derivation step in the interpretable attribution chain violates a defined logical rule, its logical support relationship is marked as problematic.

[0235] In the step of verifying whether the overall derivation process of the explainable attribution chain conforms to the physical consistency constraints defined in the inference skeleton and generating logical reliability assessment results, the aim is to ensure that the derivation process of the explainable attribution chain is not only logically self-consistent but also conforms to the physical laws of equipment operation. By comparing the causal relationships in the explainable attribution chain with the predefined physical consistency constraints in the inference skeleton, it can be determined whether the entire diagnostic process matches the physical model of the equipment. One implementation is to match each causal link in the explainable attribution chain (e.g., a fault causes an abnormal physical quantity) with the corresponding physical consistency constraint in the inference skeleton. For example, if the attribution chain indicates "bearing wear causes abnormal vibration," and the physical consistency constraint explicitly defines the physical correlation between "bearing wear" and "an increase in a specific frequency component of the vibration spectrum," then the link is considered to conform to the constraint. If there is a mismatch or violation of the constraint, it is marked as non-conforming. Another implementation is to verify it through simulation. The fault mode and propagation path described by the explainable attribution chain are input into the equipment physical model for simulation, and the simulation results are observed to see if they are consistent with the physical performance predicted by the attribution chain. If the simulation results deviate significantly from the predictions, the attribution chain is considered not to meet the physical consistency constraint.

[0236] The step of integrating the results of the coverage integrity audit and the logical reliability assessment to generate a reflective assessment report, which includes identifications of coverage gaps, logical conflict issues, and confidence scores for each item, aims to summarize the results of the two audits and assessments into a comprehensive reflective assessment report, providing a basis for decision-making in subsequent closed-loop optimization. This report not only identifies potential knowledge coverage deficiencies and reasoning logic problems during the diagnostic process but also quantifies the severity of these problems through confidence scores. One implementation approach is to design a structured report template, filling in the uncovered potential failure modes identified in the coverage integrity audit results as coverage gaps, and the logical inconsistencies or physical discrepancies identified in the logical reliability assessment results as logical conflict issues. Simultaneously, a confidence score is calculated or assigned to each identifier, for example, based on semantic similarity or logical conflict strength. Another implementation approach is to employ methods based on expert systems or machine learning models. Using the raw audit and assessment results as input, a pre-trained model automatically generates a reflective assessment report containing coverage gaps, logical conflict issues, and their confidence scores. Confidence scores can be given based on the model's assessment of the severity of each problem.

[0237] Through the above technical solutions, this application can systematically audit the coverage and completeness of diagnostic hypotheses. By comparing the fault hypothesis space to be verified in the inference skeleton with the historical fault case nodes in the associated knowledge subgraph paths, it effectively identifies potential fault modes not covered by the associated knowledge subgraph paths. This solves the coverage blind spot problem caused by the lag in knowledge base updates, ensures the comprehensiveness of diagnostic hypotheses, and avoids the omission of potential faults. Simultaneously, by verifying the logical support relationship between the evidence nodes on which the explainable attribution chain depends based on consistency markers in the structured evidence network, and verifying whether the overall derivation process of the explainable attribution chain conforms to the physical consistency constraints defined in the inference skeleton, the logical reliability of the core evidence chain can be comprehensively assessed. This solves the problem of unreliable reasoning caused by "floating" or contradictory evidence, and ensures that the diagnostic derivation process is highly consistent with the physical model of the equipment, effectively avoiding "logical defocus" or "physical decoupling." Integrating these audit and evaluation results generates a reflective evaluation report that includes identification of missing items, identification of logical conflicts, and confidence scores for each item. This provides a precise and traceable basis for subsequent closed-loop optimization instructions (such as immediate verification instructions for scheduling supplementary evidence collection, or long-term evolutionary instructions for updating the domain knowledge graph and device physical model). This greatly improves the accuracy of metacognitive reflection and the effectiveness of closed-loop optimization instructions, thereby enabling the entire intelligent diagnostic system to have stronger adaptive and self-optimizing capabilities.

[0238] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0239] Figure 9 This is a schematic diagram of the structure of a data processing and insight system based on a multimodal intelligent assistant provided in an embodiment of this application. See also... Figure 9 The system includes: The deduction and retrieval module 901 is used to deduce the fault mechanism of the equipment physical model to obtain the physical fault prior set, and to perform context retrieval on the domain knowledge graph to obtain the associated knowledge subgraph path; The skeleton generation module 902 is used to perform dual-track coupling and constraint injection on the physical fault prior set and the associated knowledge subgraph path to generate a reasoning skeleton. The reasoning skeleton defines the fault hypothesis space to be verified, the multimodal evidence requirements, and the embedded physical and spatiotemporal constraints. The verification module 903 is used to schedule multiple heterogeneous intelligent agents to perform collaborative evidence collection based on the inference skeleton, and to perform conflict detection and cross-verification based on the evidence acquired in real time during the collection process, forming a structured evidence network with consistency tags. The instruction generation module 904 is used to perform dynamic graph fusion and traceable reasoning on the structured evidence network and the associated knowledge subgraph path to generate a comprehensive credibility assessment and an interpretable attribution chain. Based on the comprehensive credibility assessment and the consistency mark, it triggers metacognitive reflection to generate a closed-loop optimization instruction. The closed-loop optimization instruction is used to perform supplementary verification or update the domain knowledge graph and the device physical model.

[0240] It should be noted that the data processing and insight system based on a multimodal intelligent assistant provided in the above embodiments is only illustrated by the division of the above functional modules when performing model early warning. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the data processing and insight system based on a multimodal intelligent assistant provided in the above embodiments and the data processing and insight method embodiments based on a multimodal intelligent assistant belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0241] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program that can be executed by a processor to perform the data processing and insight method based on a multimodal intelligent assistant in the above embodiments. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0242] In an exemplary embodiment, a computer program product or computer program is also provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to perform the above-described data processing and insight method based on a multimodal intelligent assistant.

[0243] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0244] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0245] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-modal intelligent assistant based data processing and insights method, characterized in that, The method includes: The physical model of the equipment is used to deduce the fault mechanism to obtain the physical fault prior set, and the domain knowledge graph is used to perform contextual retrieval to obtain the path of the associated knowledge subgraph; The physical fault prior set and the associated knowledge subgraph path are coupled in a dual-track manner and constraint injection is performed to generate a reasoning skeleton. The reasoning skeleton defines the fault hypothesis space to be verified, the multimodal evidence requirements, and the embedded physical and spatiotemporal constraints. Based on the inference framework, multiple heterogeneous intelligent agents are scheduled to perform collaborative evidence collection, and conflict detection and cross-validation are performed based on the evidence acquired in real time during the collection process to form a structured evidence network with consistency tags; The structured evidence network and the associated knowledge subgraph path are dynamically fused and traceable reasoned to generate a comprehensive credibility assessment and an explainable attribution chain. Based on the comprehensive credibility assessment and the consistency marker, metacognitive reflection is triggered to generate a closed-loop optimization instruction. The closed-loop optimization instruction is used to perform supplementary verification or update the domain knowledge graph and the device physical model.

2. The method according to claim 1, characterized in that, The process of performing fault mechanism deduction on the equipment physical model to obtain a set of physical fault priors, and performing contextual retrieval on the domain knowledge graph to obtain the path of the associated knowledge subgraph, includes: Based on the current equipment operating parameters and alarm signals, an initial set of fault assumptions is generated through the equipment physical model, and the theoretical physical characteristics corresponding to the initial set of fault assumptions are determined. Using the current equipment operating parameters and the alarm signal as the retrieval context, an initial associated knowledge subgraph is retrieved from the domain knowledge graph; Based on the historical experience rules and fault associations contained in the initial association knowledge subgraph, the initial fault hypothesis set is verified and filtered to generate the physical fault prior set. Based on the physical fault prior set, a structured query is constructed for the domain knowledge graph, and a deep search of the subgraph path is performed to obtain the associated knowledge subgraph path.

3. The method according to claim 1, characterized in that, The step of performing dual-track coupling and constraint injection on the physical fault prior set and the associated knowledge subgraph path to generate a reasoning skeleton includes: Semantic alignment and association matching are performed between the fault modes in the physical fault prior set and the fault nodes in the associated knowledge subgraph path to construct a mapping relationship network between fault modes and knowledge nodes. Based on the mapping relationship network, the theoretical physical features in the physical fault prior set and the historical case context in the associated knowledge subgraph path are fused and complementary to generate a unified fault hypothesis space to be verified, and the required multimodal evidence type is assigned to each fault hypothesis in the fault hypothesis space to be verified. The physical consistency constraints corresponding to the fault hypothesis space to be verified are derived from the equipment physical model deduction logic, and the spatiotemporal acquisition constraints are derived from the equipment topology of the associated knowledge subgraph path. The physical consistency constraints and the spatiotemporal acquisition constraints are used as structured metadata and associated with the corresponding fault hypotheses in the fault hypothesis space to be verified and the multimodal evidence types, respectively, to assemble and generate the inference skeleton.

4. The method according to claim 3, characterized in that, Based on the mapping relationship network, the theoretical physical characteristics in the physical fault prior set and the historical case context in the associated knowledge subgraph path are fused and complementaryly corrected to generate a unified fault hypothesis space to be verified, including: For each fault mode associated in the mapping network, obtain the corresponding theoretical physical characteristics and the set of actual observed characteristics from the historical case context; A statistical difference analysis is performed on the theoretical physical characteristics and the actual observation characteristics set to obtain the analysis results. Based on the analysis results, the parameter range of the theoretical physical characteristics is calibrated, and the missing feature dimensions are supplemented from the actual observation characteristics set to obtain the calibrated and supplemented feature descriptions. Based on the calibrated and supplemented feature descriptions, consistency verification and conflict resolution are performed on the descriptions of the same fault mode from different orbits. Unreconcilable feature conflicts are eliminated to obtain the fault modes that pass the verification and the calibrated feature descriptions. All verified fault modes and calibrated feature descriptions are integrated to form the fault hypothesis space to be verified.

5. The method according to claim 1, characterized in that, The step of scheduling multiple heterogeneous agents to perform collaborative evidence collection based on the inference framework includes: The reasoning framework is analyzed to extract the fault hypothesis to be verified and the associated multimodal evidence requirements, physical constraints and spatiotemporal acquisition constraints. Based on the multimodal evidence requirements and the spatiotemporal acquisition constraints, an atomized evidence acquisition task description is generated for each fault hypothesis, and a matching heterogeneous intelligent agent is assigned to execute each atomized evidence acquisition task description. The physical constraints and the spatiotemporal acquisition constraints are converted into acquisition instruction parameters that can be executed by the heterogeneous intelligent agent, and the task startup sequence of the heterogeneous intelligent agent is coordinated based on the logical relationship between the steps in the inference skeleton. The multiple heterogeneous intelligent agents are scheduled to execute atomic evidence collection tasks according to the collection command parameters, and the multimodal raw evidence data output by the heterogeneous intelligent agents is accompanied by a unified spatiotemporal synchronization mark.

6. The method according to claim 1, characterized in that, The process of performing conflict detection and cross-validation based on real-time acquired evidence during the collection process to form a structured evidence network with consistency markers includes: Modal analysis and spatiotemporal synchronization alignment are performed on the real-time acquired evidence to generate multiple standardized evidence nodes, which constitute a set of standardized evidence nodes. Each standardized evidence node includes evidence content, modal type, spatial location, and acquisition timestamp. Based on the modal type and content of each standardized evidence node in the standardized evidence node set, cross-modal logical consistency verification and numerical compatibility analysis are performed to identify and mark conflicting evidence node pairs that contradict each other. For the marked conflict evidence node pairs, targeted cross-validation is triggered: heterogeneous agents related to the conflict mode are scheduled, retesting is performed according to the spatial location involved in the conflict, verification evidence nodes are generated, and the verification evidence nodes are added to the standardized evidence node set. The structured evidence network is constructed by using all nodes in the standardized evidence node set as network nodes and the supporting, contradictory, or verifying relationships between nodes as edges.

7. The method according to claim 6, characterized in that, Based on the modal type and content of the standardized evidence nodes, the process of performing cross-modal logical consistency verification and numerical compatibility analysis to identify and mark conflicting evidence node pairs that contradict each other includes: Based on the spatial location and collection timestamp of the standardized evidence nodes, standardized evidence nodes of different modalities are spatiotemporally associated and paired to form a cross-modal evidence node pair set; For each node pair in the cross-modal evidence node pair set, the evidence content of the standardized evidence nodes in the node pair is parsed, and semantic-level logical consistency verification is performed based on the domain knowledge base of the fault diagnosis field. For each node pair in the cross-modal evidence node pair set, extract the numerical features of the standardized evidence nodes in the node pair, and perform numerical compatibility analysis based on the physical consistency constraints; Cross-modal evidence node pairs that are indicated as contradictory by the logical consistency verification result or as incompatible by the numerical compatibility analysis result are marked as conflicting evidence node pairs, and conflict type and conflict intensity measure are recorded for the conflicting evidence node pairs.

8. The method according to claim 1, characterized in that, The step of dynamically fusing the structured evidence network with the associated knowledge subgraph paths and performing traceable reasoning to generate a comprehensive credibility assessment and an explainable attribution chain includes: The standardized evidence nodes in the structured evidence network are used as temporary nodes and dynamically inserted into the path of the associated knowledge subgraph to construct a fusion reasoning graph; Graph structure reasoning is performed on the fused reasoning graph to identify the optimal reasoning path connecting key evidence nodes and candidate fault nodes, and the optimal reasoning path is converted into the interpretable attribution chain. Based on the node relationships in the fused reasoning graph, the consistency markers in the structured evidence network, and the physical consistency constraints in the reasoning skeleton, the comprehensive credibility assessment is determined. The comprehensive credibility assessment integrates physical consistency confidence, multimodal consistency confidence, and knowledge graph support.

9. The method according to claim 8, characterized in that, The determination of the comprehensive credibility assessment based on the node relationships in the fused inference graph, the consistency markers within the structured evidence network, and the physical consistency constraints in the inference skeleton includes: Based on the physical consistency constraints defined in the inference skeleton, the degree of physical feature matching of standardized evidence nodes in the structured evidence network is evaluated, and the physical consistency confidence is calculated. Based on the consistency markers within the structured evidence network, the collaborative support status and conflict degree among cross-modal evidence nodes are quantified, and the multimodal consistency confidence is calculated. Based on the weights of the associated edges and the path support between the standardized evidence nodes and knowledge nodes in the fusion reasoning graph, the support degree of the knowledge graph is calculated. Based on a preset weight allocation strategy, the physical consistency confidence, the multimodal consistency confidence, and the knowledge graph support are weighted and fused to generate the comprehensive credibility assessment.

10. The method according to claim 1, characterized in that, The step of triggering metacognitive reflection based on the comprehensive credibility assessment and the consistency marker to generate closed-loop optimization instructions includes: The comprehensive credibility assessment is compared with a preset credibility threshold, and the consistency markers in the structured evidence network are analyzed. Based on the comparison and analysis results, it is determined whether metacognitive reflection is triggered. When metacognitive reflection is triggered, the reasoning skeleton, structured evidence network, related knowledge subgraph paths, and interpretable attribution chains related to the current diagnostic task are obtained as reflection context data. Based on the aforementioned reflective context data, an audit of the completeness of the diagnostic hypothesis coverage and an assessment of the logical reliability of the core chain of evidence are performed, generating a reflective assessment report; Based on the reflection and evaluation report, the closed-loop optimization instructions are generated. The closed-loop optimization instructions include immediate verification instructions for scheduling supplementary evidence collection, or long-term evolution instructions for updating the domain knowledge graph and the device physical model.

Images