Knowledge graph-based industrial equipment fault root cause analysis method and system

Abstract

The application provides an industrial equipment fault root cause analysis method and system based on a knowledge graph. The method comprises the following steps: obtaining vibration data, temperature data and operation parameter data of an industrial equipment, and generating a standardized equipment fault evidence set; based on a pre-constructed industrial equipment fault knowledge graph, calculating the correlation score between a single equipment fault evidence and a candidate fault mode node; calculating the comprehensive score and coverage rate of the root cause candidate chain, and generating an industrial equipment fault root cause candidate set; determining the explanation consistency score of each root cause candidate chain to the evidence in the window according to the industrial equipment fault root cause candidate and the standardized equipment fault evidence set; and outputting the fault type, failure component and corresponding explanation chain structure in the order of the final decision score. The application more specifically solves the problems in the prior art that abnormal information is difficult to be uniformly organized, knowledge is disconnected from real-time diagnosis, candidate reasons are difficult to verify, and final conclusions are difficult to be implemented.

Description

Technical Field

[0001] This invention belongs to the field of root cause analysis of industrial equipment failures, and particularly relates to a method and system for root cause analysis of industrial equipment failures based on knowledge graphs. Background Technology

[0002] Industrial equipment in petrochemical, power, metallurgical, rail transportation, mining, and manufacturing industries operates under continuous, heavy-load, and complex switching conditions for extended periods. Its operational reliability directly impacts production continuity, equipment lifespan, and maintenance costs. With the development of online monitoring technology and industrial control systems, it is now possible to continuously acquire vibration data, temperature data, and operating parameters such as speed, load, and current on-site, and to achieve anomaly monitoring and early warning through edge computing and cloud platforms. In existing technologies, one type of solution mainly identifies abnormal states through threshold judgment, time-domain or frequency-domain feature analysis, trend analysis, and rule-based matching. This type of solution is suitable for known faults with relatively clear boundaries, but it often remains at the level of identifying anomalies in a single monitoring quantity or a single component, making it difficult to handle industrial scenarios where components are strongly coupled, anomaly propagation chains are long, and apparent phenomena are intertwined with the root cause of the fault. Another type of solution emphasizes data-driven fault classification, health assessment, or life prediction, and establishes fault identification models by training on historical samples. However, industrial sites generally have problems such as a small number of real fault samples, incomplete labels, large differences in equipment models, and frequent changes in operating conditions, resulting in significant shortcomings of this type of solution in terms of cross-equipment adaptation, root cause tracing, and engineering interpretation. In recent years, solutions utilizing knowledge graphs to organize equipment structure, component relationships, anomalies, and failure modes have gradually emerged. These solutions can, to some extent, compensate for the shortcomings of pure data methods in expressing equipment mechanisms. However, most existing related technologies are still limited to knowledge retrieval, case querying, or simple node matching. They lack a unified processing chain for the task of root cause analysis of industrial equipment. In particular, there is a lack of a technical approach that can first organize multi-source anomalies under the same analysis window into standardized equipment failure evidence with component location and anomaly combination semantics, then use equipment structure relationships, anomaly phenomenon relationships, and failure mode relationships to map this evidence into a local association structure oriented towards the current failure scenario, and further form a comparable, sortable, and interpretable root cause candidate chain on this local association structure.

[0003] Therefore, in industrial equipment scenarios involving multiple components and concurrent anomalies, existing technologies are still prone to problems such as scattered anomaly information, lack of unified expression among evidence, inability of knowledge graphs to truly participate in real-time root cause reasoning, difficulty in closed-loop verification of candidate causes and on-site anomalies, and insufficient interpretability of final diagnostic results. It is difficult to simultaneously ensure the accuracy of root cause localization, the completeness of the analysis chain, and the availability of operation and maintenance decisions. Summary of the Invention

[0004] The purpose of this invention is to propose a method and system for root cause analysis of industrial equipment failures based on knowledge graphs, thereby solving the above-mentioned problems.

[0005] To achieve the above objectives, a knowledge graph-based root cause analysis method for industrial equipment failures is provided in a first aspect of the present invention, the method comprising the following steps: S1. Acquire vibration data, temperature data, and operating parameter data of industrial equipment, and simultaneously organize them to generate a standardized equipment failure evidence set; wherein, each piece of equipment failure evidence in the standardized equipment failure evidence set includes analysis window identifier, evidence strength, component location, and abnormality type combination information; S2. Based on the pre-built industrial equipment fault knowledge graph, each piece of equipment fault evidence in the standardized equipment fault evidence set is mapped to component nodes, abnormal phenomenon nodes, and fault mode nodes. The association score between a single piece of equipment fault evidence and a candidate fault mode node is calculated. The path is then filtered by combining the retention score of the corresponding candidate path and the convergence degree of multiple evidences to form an evidence association subgraph. S3. Based on the evidence association subgraph, multiple paths are merged around the same fault mode node to form a root cause candidate chain. The comprehensive score and coverage of the root cause candidate chain are calculated. The comprehensive score is obtained by combining the mean of the path scores with the structural closure enhancement coefficient and the abnormal conflict suppression coefficient. The coverage is obtained by the proportion of the number of evidence pieces covered by the candidate chain to the total number of evidence pieces in the window. A root cause candidate set for industrial equipment failure is generated. S4. Based on the industrial equipment failure root cause candidates and the standardized equipment failure evidence set, determine the interpretability score of each root cause candidate chain for the evidence within the window; combine the comprehensive score, coverage and interpretability score to obtain the final decision score, and output the failure type, failed component and corresponding interpretability chain structure according to the final decision score as the result of industrial equipment failure root cause analysis.

[0006] Furthermore, the steps of the synchronization and sorting are as follows: Based on the vibration data, temperature data, and operating parameter data, vibration impact features, temperature change features, and operating condition deviation features are extracted according to the analysis window, and then normalized and weighted fused to generate a standardized equipment fault evidence set. Within each analysis window, the vibration data is used to extract the vibration impact characteristics, specifically the difference between the vibration peak value and the vibration mean value within that window, which is used to characterize local mechanical impact. Temperature data is used to extract temperature change features, specifically the difference between the current window temperature and the previous window temperature, which is used to characterize short-term temperature rise. The operating parameter data is used to extract the operating parameter offset features, specifically the degree of deviation of the speed, load or current from the center value of the relatively stable operating range within the current window, which is used to characterize the operating condition disturbance level of this window.

[0007] Furthermore, the weights of the weighted fusion are preset according to the device type and component fault sensitivity; The strength of the evidence is the weighted sum of the vibration and impact characteristics, temperature change characteristics, and operating condition deviation characteristics with their corresponding weights; The standardized equipment failure evidence set includes analysis window identifiers, component locations, anomaly combination semantics, and evidence strength. The anomaly type combination information is the corresponding basic anomaly phenomenon, which is a joint interpretation of vibration and shock, temperature change and operating condition deviation.

[0008] Furthermore, based on the pre-built industrial equipment fault knowledge graph, each piece of equipment fault evidence in the standardized equipment fault evidence set is mapped to component nodes, anomaly nodes, and fault mode nodes, specifically as follows: For each piece of standardized equipment failure evidence, first locate the corresponding component node in the component index table of the industrial equipment failure knowledge graph based on the component location, and then complete the combination disassembly and abnormal phenomenon node location in the abnormal phenomenon word list based on the abnormal type combination information. Extending from component nodes along component connection relationships to first-order adjacent components, retrieving possible corresponding fault mode nodes from the abnormal phenomenon nodes along fault cause relationships, and generating a candidate node set around the current equipment fault evidence.

[0009] Furthermore, the association score is calculated and generated based on the strength of evidence of equipment failure, the anomaly semantic matching coefficient, and the component structure consistency coefficient; Among them, the abnormal semantic matching coefficient is determined by the degree of keyword matching between the abnormal combination in the evidence and the incident abnormal phenomenon node of the candidate fault mode node, and the component structure consistency coefficient is determined by the connection level of the component position in the evidence and the component set associated with the candidate fault mode node in the equipment structure diagram.

[0010] Furthermore, the process of combining the retention scores of the corresponding candidate paths with the convergence degree of multiple evidences to filter paths and form an evidence association subgraph is as follows: The retention score is calculated and generated based on the association score of the target failure mode node pointed to by the current candidate path, the evidence convergence coefficient, and the hierarchical length of the current path. The evidence convergence coefficient is the ratio of the number of device failure evidence items that are under the same analysis window identifier as the current candidate path and point to the same target failure mode node to the total number of device failure evidence items under that analysis window identifier. The level length of the current path is the number of edges traversed from the starting component node to the target failure mode node. Based on preset thresholds and path retention scores, high-scoring nodes and paths are selected to form an evidence association subgraph.

[0011] Furthermore, the nodes of the evidence association subgraph include component nodes, anomaly nodes, and fault mode nodes related to the input evidence set, and the edges include component connection relationships, component performance relationships, and fault cause relationships; wherein, each fault mode node is accompanied by association score and path score information contributed by several equipment fault evidences, and each path contains equipment fault evidence.

[0012] Furthermore, the comprehensive score is calculated and generated based on the mean retention score of each path in the root cause candidate chain, the structural closure enhancement coefficient, and the abnormal conflict suppression coefficient. The structural closure enhancement coefficient is determined based on the proximity of each component node in the equipment structure diagram, and the abnormal conflict suppression coefficient is determined based on whether the abnormal type combinations of each piece of evidence in the chain support each other. The coverage rate represents the coverage of the root cause candidate chain to the device failure evidence set under the same analysis window identifier. It is calculated based on the number of device failure evidence pieces actually covered by the candidate root cause chain under the analysis window identifier and the total number of device failure evidence pieces under the same analysis window identifier.

[0013] Furthermore, the interpretation consistency score is calculated based on the sum of the evidence strengths of any device failure evidence under the same analysis window identifier as the current candidate root cause chain and the interpretation identifier of the corresponding device failure evidence relative to the current candidate root cause chain. The sum of the evidence strengths of any device failure evidence under the same analysis window identifier as the current candidate root cause chain is the sum of the evidence strengths of all device failure evidence under the same analysis window identifier; Specifically, when the equipment failure evidence can be explained by at least one branch path in the current candidate root cause chain, the explanation identifier of the corresponding equipment failure evidence relative to the current candidate root cause chain is set to 1; otherwise, it is set to 0.

[0014] A second aspect of the present invention provides a knowledge graph-based industrial equipment failure root cause analysis system, the system comprising: The evidence generation module is used to acquire vibration data, temperature data, and operating parameter data of industrial equipment, and synchronously organize them to generate a standardized equipment failure evidence set. Each piece of equipment failure evidence in the standardized equipment failure evidence set includes analysis window identifier, evidence strength, component location, and abnormality type combination information. The evidence generation module is used to map each piece of equipment failure evidence in the standardized equipment failure evidence set to component nodes, abnormal phenomenon nodes, and failure mode nodes based on a pre-built industrial equipment failure knowledge graph. It calculates the association score between a single piece of equipment failure evidence and a candidate failure mode node, and combines the retention score of the corresponding candidate path and the convergence degree of multiple evidences to filter the path and form an evidence association subgraph. The candidate chain generation module is used to merge multiple paths around the same fault mode node to form a root cause candidate chain based on the evidence association subgraph, calculate the comprehensive score and coverage of the root cause candidate chain, the comprehensive score is obtained by combining the mean of the path scores with the structural closure enhancement coefficient and the abnormal conflict suppression coefficient, and the coverage is obtained by the proportion of the number of evidence pieces covered by the candidate chain to the total number of evidence pieces in the window, and generate a root cause candidate set for industrial equipment faults. The root cause output module is used to determine the interpretability score of each root cause candidate chain for the evidence within the window based on the industrial equipment failure root cause candidates and the standardized equipment failure evidence set; combine the comprehensive score, coverage and interpretability score to obtain the final decision score, and output the failure type, failed component and corresponding interpretability chain structure according to the final decision score as the root cause analysis result of industrial equipment failure.

[0015] The beneficial technical effects of the present invention are at least as follows: This invention establishes a technical chain that converges layer by layer from on-site anomalies to root cause conclusions, centered around "root cause analysis of industrial equipment failures based on knowledge graphs". Its core is not to introduce a certain identification model or reasoning algorithm in isolation, but to close the industrial equipment operation data, knowledge organization methods and root cause determination logic within the same analysis framework. Specifically, this invention first transforms vibration data, temperature data, and operating parameters into standardized equipment failure evidence with semantics of evidence strength, component location, and abnormal combination, all within a single analysis window. This allows anomalies from different monitoring channels to enter subsequent analysis in a unified form. Based on this, relying on a pre-constructed industrial equipment failure knowledge graph, the standardized equipment failure evidence is associated with component entities, anomaly entities, and failure mode entities. An evidence association subgraph is formed by combining component connection relationships, component performance relationships, and failure causal relationships, transforming the graph from a static knowledge carrier into a local reasoning structure directly carrying the current failure scenario. Furthermore, around multiple paths converging on the same failure mode node in the evidence association subgraph, a candidate chain of industrial equipment failure root causes is constructed, capable of simultaneously explaining multiple component locations and multiple anomalies. The candidate chains are compared through chain-level scoring, coverage, and interpretive consistency, elevating root cause analysis from single-node judgment to a comprehensive judgment based on the evidence chain. Finally, the explanatory power of the candidate chain for standardized equipment failure evidence is combined with the structural rationality of the candidate chain itself to output the failure type, the failed component, and its corresponding interpretive chain structure, thereby achieving traceable analysis from abnormal manifestations to root causes.

[0016] Through the above technical solutions, the present invention can unify multi-source anomalies into computable equipment failure evidence, transform knowledge graphs into reasoning carriers directly related to the current anomaly scenario, and improve the accuracy of root cause localization and the engineering interpretability of the results through bidirectional verification between candidate chains and evidence. This more effectively solves the problems in the prior art, such as the difficulty in unifying the organization of anomaly information, the disconnect between knowledge and real-time diagnosis, the difficulty in verifying candidate causes, and the difficulty in implementing final conclusions. Attached Figure Description

[0017] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0018] Figure 1 This is a flowchart of the knowledge graph-based root cause analysis method for industrial equipment failures according to the present invention. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] like Figure 1 As shown in the embodiment of the present invention, the method for root cause analysis of industrial equipment failures based on knowledge graphs includes: S1. Acquire vibration data, temperature data, and operating parameter data of industrial equipment and organize them synchronously to generate a standardized equipment fault evidence set; wherein, each piece of equipment fault evidence in the standardized equipment fault evidence set includes analysis window identifier, evidence strength, component location, and abnormality type combination information.

[0021] Specifically, the vibration data, temperature data, and operating parameter data of industrial equipment are first synchronized and organized, and then transformed into a standardized set of equipment fault evidence.

[0022] Understandably, in actual deployment, vibration data is continuously collected by accelerometers installed on bearing housings, gearbox housings, or motor end covers; temperature data is periodically collected by temperature sensors located near the outer ring of the bearing, lubrication points, or hot spots in the housing; and operating parameter data is read from the communication interface of the PLC, DCS, or frequency converter, preferably selecting state variables that can characterize changes in operating conditions, such as speed, load, and current. After collection, the raw data is not directly compared point by point. Instead, the entire time period is divided into continuous analysis windows, using the control system's state refresh cycle as a unified time reference. This is because operating parameters naturally correspond to the equipment's operating condition switching process, allowing high-frequency vibration segments and low-frequency temperature changes to be grouped into the same operating segment. Taking a centrifugal pump as an example, if the control system refreshes the speed and load status at a fixed cycle, this cycle is used as an analysis window. Within each window, the corresponding vibration sequence, temperature sampling value, and operating parameter records are extracted, thus forming a joint observation unit of "the same time period, the same operating condition, and the same component location."

[0023] Preferably, within each analysis window, vibration data is used to extract vibration impact characteristics, specifically the difference between the vibration peak value and the vibration mean within that window, to characterize local mechanical impact; temperature data is used to extract temperature change characteristics, specifically the difference between the current window temperature and the previous window temperature, to characterize short-term temperature rise; and operating parameter data is used to extract operating parameter offset characteristics, specifically the degree of deviation of the rotational speed, load, or current from the center value of the stable operating range within the current window, to characterize the level of operating condition disturbance in this window. The sources of these three types of characteristics are not arbitrarily set. The difference between the vibration peak value and the mean comes from the classic statistical characterization approach for impact faults in mechanical fault diagnosis. Its original basis is that mechanical impact causes spikes in the time-domain signal; the more pronounced the spike, the greater the difference between the peak value and the mean. The temperature difference comes from the first-order difference in discrete-time series analysis, used to describe the rate of state change; and the operating parameter offset comes from deviation analysis in process control, used to characterize the degree of deviation of the current operating condition from the stable operating condition. After obtaining these three basic quantities, they are not directly added together. Instead, a historical stable operating segment of the equipment is used as a reference to map each basic quantity to a unified scale. The mapping method employs the classic linear normalization approach: for each feature, the common value range under stable operating conditions is pre-calculated, and then the feature values ​​of the current window are converted into dimensionless results according to this range. After this processing, vibration and shock, temperature changes, and operating condition deviations all fall into the same comparison space, and subsequent fusion has clear engineering significance.

[0024] The strength of evidence for standardized equipment failure is based on a linear weighted comprehensive evaluation model from mathematics. The original form of this model involves a linear combination of multiple dimensionless indicators with weights to form a comprehensive score, commonly used for multi-indicator comprehensive decision-making. This application does not directly apply the original form, but makes two restrictive modifications to suit the root cause analysis scenario of industrial equipment: First, the indicators involved in the combination are no longer general evaluation indicators, but consist of vibration and shock characteristics, temperature change characteristics, and operating condition deviation characteristics, which respectively correspond to the three observation surfaces most directly related to the root cause of failure: mechanical shock, thermal anomaly, and operating condition disturbance; Second, the weights are not fixed universal values, but are preset according to equipment type and component failure sensitivity, making the strength of evidence more consistent with the abnormal expression patterns of specific equipment. Based on this, the strength of evidence for equipment failure is calculated using the following formula: ; In the formula, The strength of evidence for equipment failure corresponding to the current analysis window and the current component location; The normalized result of the vibration and shock characteristics is obtained by combining the difference between the vibration peak value and the vibration mean value within the window with the historical stable operating range. The normalized result of the temperature change characteristics is obtained by combining the difference between the current window temperature and the previous window temperature with the historical stable operating range. The normalized result of the operating parameter offset characteristics is obtained by combining the deviation of the speed, load or current from the center value of the stable operating range within the current window with the historical stable operating range. , , The fusion weights corresponding to the three types of features are derived from the equipment type configuration table or the statistical results of historical maintenance experience, and are preset according to equipment type and component fault sensitivity. The derivation logic of this formula is to first extract three basic observations separately, then transform the three basic observations to the same scale, and finally perform linear fusion according to the weights that match the equipment mechanism. Therefore, it is not an isolated formula, but the result of a continuous processing chain of "windowed sampling - feature extraction - normalization - weighted fusion".

[0025] In practical application, the installation location of each sensor is first linked to the location of a component in the equipment structure. For example, an accelerometer mounted on the front bearing housing is linked to the "spindle front bearing," and a temperature sensor mounted on the gearbox wall is linked to the "gear meshing adjacent area." Then, the strength of evidence for each component location is calculated in windows. For instance, the difference between the peak and mean vibration values ​​at the spindle front bearing location within a certain window is obtained after historical interval conversion. The temperature difference is obtained after conversion. The running parameter offset is obtained after conversion. For this type of rotating equipment, the configuration is as follows: , , After substituting, we get By combining the component location corresponding to this window with the semantic combination of anomalies, a complete piece of evidence for equipment failure can be formed, such as "the front bearing of the spindle exhibits increased impact and temperature rise under rated operating conditions, with an evidence strength of 0.62." If the gearbox area in another window is obtained... , , Similarly, if a high percentage of operating condition disturbances is obtained after substituting the values, the equipment fault evidence formed in this window will be marked as "gearbox area anomaly significantly affected by operating condition disturbances". This results not in scattered raw signals or single sensor alarms, but in equipment fault evidence that already possesses component-specificity, anomaly combination semantics, and intensity quantification results.

[0026] The final output is a standardized set of equipment failure evidence. Each piece of equipment failure evidence in this set consists of four parts: first, the component location, derived from the correspondence between sensor installation points and the equipment structure table; second, the analysis window identifier, derived from the window segmentation results under a unified time base; third, anomaly type combination information, derived from the joint interpretation of vibration and shock, temperature changes, and operating condition deviations; and fourth, the strength of the evidence. , derived from the aforementioned weighted fusion calculation.

[0027] S2. Based on the pre-constructed industrial equipment fault knowledge graph, each piece of equipment fault evidence in the standardized equipment fault evidence set is mapped to component nodes, abnormal phenomenon nodes, and fault mode nodes. The association score between a single piece of equipment fault evidence and a candidate fault mode node is calculated. The path is then filtered by combining the retention score of the corresponding candidate path and the convergence degree of multiple evidences to form an evidence association subgraph.

[0028] Specifically, obtain a standardized set of evidence for equipment failures.

[0029] Understandably, each piece of equipment failure evidence in this set includes an analysis window identifier and evidence strength. This step focuses on how equipment failure evidence under the same analysis window identifier enters the industrial equipment failure knowledge graph and forms a reasonable local structure. The industrial equipment failure knowledge graph is pre-formed from equipment structural design data, component assembly relationships, typical failure mechanism data, and historical maintenance cases. The graph contains at least three types of nodes: component entities, abnormal phenomenon entities, and failure mode entities, as well as three types of edges: component connection relationships, component performance relationships, and failure cause relationships. Master data unification has been completed during the graph construction phase. For example, "spindle front bearing," "front bearing," and "bearing position 1" are merged into the same component entity, and "impact enhancement," "impact mutation," and "increased spikes" are merged into the same abnormal phenomenon entity. Therefore, the component location and abnormal type combination information output in step one can be directly used as the mapping entry point for this step. For each piece of equipment failure evidence, the corresponding component node is first located in the component index table of the graph based on the component location. Then, the combination and disassembly of the abnormal phenomenon terminology are completed based on the abnormal type combination information. For example, when the equipment failure evidence output in Step 1 is described as "the front bearing of the spindle experiences increased impact and temperature rise under rated operating conditions, with an evidence strength of 0.62," the component location "front bearing of the spindle" is directly located to the component node in the graph. The anomaly type combination information is decomposed into two basic anomalies: "increased vibration and impact" and "temperature rise," which are then located to their respective anomaly phenomenon nodes. Next, the process extends from the component node along the component connection relationship to first-order adjacent components, and from the anomaly phenomenon node along the fault causation relationship, it retrieves possible corresponding fault mode nodes, thus forming a candidate node set surrounding the current equipment failure evidence. This process is specifically targeted at industrial equipment scenarios because equipment failure propagation is naturally constrained by structural proximity relationships. The same anomaly falling on different component locations often points to different root causes. Similarly, when multiple anomalies act on the same component, they significantly improve the interpretability of certain fault modes. Therefore, the mapping process simultaneously retains the two core information elements, "location" and "anomaly combination," from the evidence in Step 1, ensuring that the evidence maintains its engineering semantic integrity after entering the graph.

[0030] Understandably, the strength of the correlation between equipment failure evidence and candidate failure mode nodes is calculated using an improved weighted similarity evaluation formula. Its original source is the linear weighted similarity model in mathematics, which uses a weighted sum of multiple normalized indices to represent the overall matching degree between objects. This step involves two layers of deduction based on this model. The first layer of deduction replaces the "feature components" in the traditional similarity evaluation with two quantities more suitable for root cause analysis of industrial equipment: the anomaly semantic matching coefficient and the component structure consistency coefficient. The anomaly semantic matching coefficient reflects the degree of correspondence between the anomaly type combination information in step one and the set of anomalies corresponding to the current candidate failure mode node, while the component structure consistency coefficient reflects the proximity of the component location in step one to the set of components associated with the current candidate failure mode node within the equipment structure. The second layer of deduction involves adjusting the strength of the evidence obtained in step one. Introduced as a holistic coefficient, the significance of on-site anomalies directly participates in the mapping of failure mode nodes, thus unifying "the strength of the evidence itself" and "whether the evidence matches the target failure mode" within the same computational framework. Therefore, the correlation score is written as: ; in, The score represents the correlation between a single piece of equipment failure evidence and a candidate failure mode node. The strength of the evidence of equipment failure output in step one; The anomaly semantic matching coefficient is derived from the comparison between the anomaly type combination information in the equipment fault evidence and the standard descriptions of all incident anomaly nodes of the current candidate fault mode node. In actual calculation, the keyword set after decomposing the anomaly combination can be compared with the standard description keyword set of these anomaly nodes, and the one with the highest matching degree can be taken as the coefficient. ; The component structure consistency coefficient is derived from the connection level of the component location in step one and the component set associated with the current candidate fault mode node in the device structure diagram. If the component corresponding to the current device fault evidence is the same component as the component associated with the candidate fault mode node, a higher value is taken. If they are first-order adjacent components, the coefficient is reduced by a preset coefficient. If they are second-order adjacent components, the coefficient is further reduced, and the maximum consistency value is taken as the coefficient. ; This is a balancing coefficient, derived from the device type configuration table, used to adjust the weight distribution between abnormal semantic matching and structural consistency. Because... , , , All are dimensionless quantities, therefore both sides of the formula remain consistent. The derivation logic can be expressed as follows: First, obtain the on-site anomaly intensity from step one. Then, the correspondence between the current equipment fault evidence anomaly description and the corresponding set of anomalies for the current candidate fault mode node is obtained. Then, the proximity relationship between the location of the current equipment fault evidence component and the set of components associated with the current candidate fault mode node is obtained. Finally, they are combined in a linear weighted manner. and Then use The overall matching degree is modulated to obtain the final association score. Taking evidence of equipment malfunction as an example, if step one has already provided... The anomalous combination of this evidence is "increased impact accompanied by temperature rise," and the current candidate fault mode node is "bearing wear." A comparison was made between the incident anomaly node and the keywords "increased bearing impact accompanied by localized temperature rise," and their order was consistent. The corresponding component location is the same component associated with the candidate fault mode node, thus obtaining... Configuration for this type of rotating equipment Substituting, we can get If the same equipment failure evidence is used to map another candidate failure mode node, "coupling misalignment," then due to the low semantic fit of the anomaly and the fact that the components are only adjacent in structure, it can be selected as the correct choice. , After substituting, we get This clearly shows that the priority of the same piece of equipment failure evidence has been distinguished across different candidate failure mode nodes.

[0031] Furthermore, after obtaining the association scores between each piece of equipment failure evidence and each candidate node, this step further constructs local association paths and calculates retention scores for these paths to form the final evidence association subgraph. The original source of the path score is the path decay model in graph theory, the basic idea of ​​which is that the longer the path, the lower the credibility of the information during propagation. This step incorporates the "evidence convergence" factor from the industrial equipment root cause analysis scenario. This is because, in actual operation and maintenance scenarios, if a failure mode can simultaneously explain multiple pieces of equipment failure evidence within the same analysis window, its credibility is usually higher than that of a failure mode that can only explain a single piece of evidence. Therefore, this step introduces an evidence convergence coefficient on top of the path decay term to obtain the path score: ; in, This represents the retention score for a candidate path. The association score for the target failure mode node pointed to by the current candidate path; The evidence convergence coefficient represents the proportion of the number of device failure evidence items that are under the same analysis window identifier as the current candidate path and point to the same target failure mode node to the total number of device failure evidence items under that analysis window identifier. The hierarchy length of this path is determined by the graph traversal result, i.e., the number of edges traversed from the starting component node to the target failure mode node. Because... , and Since all calculations are performed in a dimensionless form, this formula remains consistent. There is a clear logical relationship between this formula and the previous one: the previous formula first addresses the question of whether a single piece of equipment fault evidence matches a single candidate node, thus obtaining a local mapping score. ; This formula is then used in Based on this, and combining path length and the degree of convergence of evidence within the same window in the graph structure, the problem of "which path is more suitable as the retained interpretation path for evidence of current device failure" is addressed. (Continuing with the above...) Taking a candidate node as an example, if the node is passed through a path of length... If the path connects to the "bearing wear" failure mode, and there are three pieces of equipment failure evidence within the same analysis window, two of which ultimately converge to this failure mode, then... Substituting, we can get If the target of the other path is "coupling misalignment", the corresponding The path length is If there is only one piece of evidence supporting the claim, then Substituting into This indicates that the "bearing wear" path possesses a higher single-evidence matching degree, a shorter structural propagation path, and a stronger multi-evidence convergence degree, therefore it should be prioritized. Statistical analysis of debugging records across several windows typically reveals that the same failure mode receives higher accuracy within the actual failure occurrence window. Paths related to operating condition disturbances but not constituting the root cause are often longer and have lower convergence. The value shows a natural decrease, and this result can be used as a basis for subsequent parameter tuning.

[0032] After completing the above calculations, all high-scoring nodes and paths corresponding to equipment failure evidence are summarized according to the analysis window to form an evidence association subgraph (where the remaining paths are sorted from high to low by the retained score, only paths with scores higher than a preset threshold are retained, and the number of paths retained under each candidate failure mode node does not exceed the upper limit set by the system). The nodes in this subgraph include component nodes, abnormal phenomenon nodes, and failure mode nodes related to the input evidence set, and the edges include component connection relationships, component performance relationships, and failure cause relationships; where each failure mode node is accompanied by association score and path score information contributed by several pieces of equipment failure evidence, and each path can be traced back to the specific equipment failure evidence in step one. Taking a complete window as an example, if the equipment failure evidence of "increased impact of the spindle front bearing accompanied by temperature rise" and the path of "bearing wear" are obtained... Evidence of equipment failure, specifically "minor current disturbance in the gearbox area," was found to be related to the "coupling misalignment" path. In this case, the evidence association subgraph formed by the window will prioritize retaining the nodes and edges corresponding to the former, while keeping the latter as a secondary explanation path on the periphery of the subgraph. The resulting graph structure is no longer a simple subset of the knowledge graph, but a local reasoning foundation driven by the equipment failure evidence from step one, and filtered according to the constraints of industrial equipment structure and the convergence rules of multiple evidences.

[0033] S3. Based on the evidence association subgraph, multiple paths are merged around the same fault mode node to form a root cause candidate chain. The comprehensive score and coverage of the root cause candidate chain are calculated. The comprehensive score is obtained by combining the mean of the path scores with the structural closure enhancement coefficient and the abnormal conflict suppression coefficient. The coverage is obtained by the proportion of the number of evidence items covered by the candidate chain to the total number of evidence items in the window. A root cause candidate set for industrial equipment faults is generated.

[0034] Specifically, an evidence association subgraph is obtained as the processing object. Each failure mode node in this evidence association subgraph has been connected to several specific paths through step two, and each path already has a path score. Furthermore, these paths can be traced back to the specific equipment failure evidence in Step One. Therefore, the specific task of this step is not to re-find the failure mode, but to merge the multiple highly relevant paths already screened in Step Two on the evidence association subgraph according to "whether the same failure mode can continuously explain multiple component locations, multiple anomalies, and multiple pieces of equipment failure evidence within the same analysis window," thereby forming a candidate chain of industrial equipment failure root causes. In actual execution, each failure mode node is first used as the central node, and all candidate paths terminating at that failure mode node are collected; then, these paths are grouped according to the component node origin, resulting in a multi-branch chain structure unfolding around the same failure mode. The object obtained in this way is no longer a single path, but a candidate root cause chain composed of multiple branches of "component node - anomaly node - failure mode node". For industrial equipment scenarios, this organizational method has direct engineering significance because real faults are usually manifested simultaneously through adjacent and coupled components. For example, bearing wear not only causes increased vibration and impact at the bearing location, but also causes temperature rise at the adjacent housing location, and may cause slight load fluctuations on the drive side. Therefore, using a multi-branch chain rather than a single path to carry the root cause explanation is more in line with the objective laws of equipment fault propagation.

[0035] Furthermore, after the candidate root cause chain is formed, its overall credibility needs to be calculated. This calculation is based on the weighted average model in statistics and the path aggregation concept in graph theory. Specifically, the quality of each path within the chain is first summarized, and then the summarized results are corrected based on the structural relationships and anomaly consistency within the chain. To make this calculation more applicable to root cause analysis of industrial equipment, this step adds two correction factors specific to the equipment scenario to the path average quality: one is the structural closure enhancement coefficient, used to reflect whether the component nodes in the same candidate root cause chain form a compact distribution around the same functional unit; the other is the anomaly conflict suppression coefficient, used to reflect whether there are obvious mutually exclusive anomalies among the fault evidence of different equipment within the chain. This modification is because the industrial equipment root cause chain not only needs a "high path score," but also needs to be "structurally cohesive" and "anomaly plausible." Based on this, the comprehensive score of the candidate root cause chain is calculated using the following formula: ; In the formula, This represents the overall score of a single candidate root cause chain; This indicates the first [number] in the candidate root cause chain. The path score for each path is directly inherited from the evidence association subgraph in step two; This indicates the number of paths contained in the candidate root cause chain, which is determined by the number of paths obtained by merging around nodes with the same failure mode; The structural closure enhancement coefficient is calculated based on the proximity of each component node in the candidate root cause chain within the device structure diagram. In actual calculation, the shortest connection level between each pair of these component nodes is first counted, and then the shorter the level and the more concentrated the distribution are mapped to a higher value. This represents the anomaly conflict suppression coefficient, calculated by whether the combinations of anomaly types among the equipment failure evidence corresponding to each path in the candidate root cause chain mutually support each other. If multiple pieces of equipment failure evidence point to the same type of mechanical shock accompanied by thermal anomalies, the value is lower; if some evidence is closer to mechanical shock-type failures and others are closer to pure operating condition disturbances, the value is higher. In this formula, It is the traditional average path quality term, derived from the weighted average model; It is a structural enhancement term added on top of the average path quality, used to reflect the local closed characteristics of the equipment structure; This is a conflict suppression term used to reflect the consistency of anomalous expressions. Because... , , All calculations use a dimensionless scoring method, thus maintaining consistency throughout the formula. The derivation process can be explicitly written as follows: First, obtain the scores for each path from step two. Summing the paths under the same failure mode and dividing by the number of paths. The basic quality of the candidate chain is obtained; then, enhancement terms are introduced based on the compactness of the relevant components within the chain in the equipment structure. Finally, suppression terms are introduced based on the consistency or conflict relationships between combinations of anomaly types within the chain. This leads to the chain-level comprehensive score. Taking a centrifugal pump as an example, if the candidate root cause chain for "bearing wear" contains three paths, step two provides the following: , , The basic mass is If these three paths correspond to the spindle front bearing, the bearing adjacent housing, and the drive-side load disturbance, respectively, and all three are distributed around the same support unit, then it can be assumed that... If the corresponding evidence of equipment failure all presents as a mutually supporting anomalous combination of increased shock, temperature rise, and slight load disturbance, then let After substituting, we get Next, examine the candidate root cause chain for "coupling misalignment." If it contains only two paths, and step two provides the... , The basic quality is If the component nodes corresponding to these two paths are scattered, it can be set as follows: If there are certain conflicts among their abnormal combinations, it can be made After substituting, we get As can be seen from the calculations, the quality of the path within the chain, the compactness of the structure, and the consistency of anomalies jointly determine the ranking of candidate root cause chains. This is consistent with the judgment logic that field engineers usually prioritize whether "multiple anomalies can converge to explain the same root cause".

[0036] Furthermore, after obtaining the comprehensive score of the candidate root cause chain... Next, this step continues by performing coverage integrity calculations on the candidate chains. This calculation is based on the concept of set coverage, which evaluates the degree to which an explanatory object covers the input set of evidence. In industrial equipment scenarios, if a candidate root cause chain can cover more equipment failure evidence within the same analysis window, and this equipment failure evidence is distributed across different but structurally related component locations, then that candidate root cause chain is more qualified as input for the subsequent final diagnosis. Therefore, this step, in the comprehensive scoring... Based on this, a coverage evaluation is introduced. The proportion of equipment failure evidence covered by a candidate chain to the total number of equipment failure evidence in the current analysis window is calculated, and this proportion serves as an important basis for retaining candidate chains. The coverage rate is calculated using the following formula: ; In the formula, This indicates the coverage of the candidate root cause chain to the equipment failure evidence set under the same analysis window identifier; This indicates the number of device failure evidence entries actually covered by the candidate root cause chain under the identifier of this analysis window. This value is obtained by checking the source of device failure evidence in step one traced back to each path in the candidate chain. This represents the total number of equipment failure evidence entries under the same analysis window identifier, calculated from the number of evidence entries with the same analysis window identifier in the standardized equipment failure evidence set output in Step 1. This formula originates from the classic definition of set coverage, and its derivation in this step is to specifically replace "elements" with "equipment failure evidence entries," thereby quantifying the explanatory power of the candidate chain for multi-source anomalies. Because... and All are counts, therefore This is a dimensionless ratio. Continuing with the centrifugal pump example above, if there are a total of [number missing] pumps in the current analysis window... There are 3 pieces of evidence of equipment failure, among which the candidate root cause chain of "bearing wear" actually covers 3. If the candidate root cause chain for "coupling misalignment" covers only one path, then... In actual processing, candidate root cause chains are arranged according to... Sort by high to low and then by coverage As a retention criterion after sorting: when two candidate root cause chains... When close, prioritize preserving coverage. Higher candidate root cause chains; when a certain candidate root cause chain... Higher but When the value is low, it is marked as a local candidate chain to help explain the source of a local anomaly; when a candidate root cause chain has a high value, it is marked as a local candidate chain. and higher If the path is not found in the target fault mode node, it is included as a priority candidate chain in the industrial equipment failure root cause candidate chain set. Therefore, the output of this step is no longer a single fault mode node or several independent paths, but a set of industrial equipment failure root cause candidate chains selected based on comprehensive scoring and coverage capability. Each candidate chain in this set consists of the target fault mode node, several high-confidence paths, and the component nodes and anomalous phenomenon nodes associated with these paths. Furthermore, each path still retains the path from step two. and At the same time, it can continue to trace back to the specific equipment failure evidence in step one.

[0037] S4. Based on the industrial equipment failure root cause candidates and the standardized equipment failure evidence set, determine the interpretability score of each root cause candidate chain for the evidence within the window; combine the comprehensive score, coverage and interpretability score to obtain the final decision score, and output the failure type, failed component and corresponding interpretability chain structure according to the final decision score as the result of industrial equipment failure root cause analysis.

[0038] Specifically, the final determination is carried out using the candidate chain set of industrial equipment failure root causes generated in step three and the standardized equipment failure evidence set generated in step one as inputs. In the candidate chain set, each candidate root cause chain already contains the candidate failure mode, the corresponding multi-branch path structure, and the comprehensive score calculated in step three. and coverage The standardized equipment failure evidence set includes evidence strength for each piece of equipment failure evidence. The key focus of this step is to elevate the requirement of "the candidate chain being structurally valid" to "the candidate chain being able to provide a complete and consistent explanation of the main anomalies within the current window," thereby forming the final root cause analysis result for industrial equipment failures. In practice, each candidate root cause chain in the candidate chain set is first expanded, its branch paths are read, and the corresponding component nodes and anomaly nodes are read. Then, the standardized equipment failure evidence set is checked to see if each piece of equipment failure evidence within the current window can be explained by the candidate chain. The establishment of the explanation relationship does not rely on additional models but on the unified semantics formed in the previous steps: if the component location in the equipment failure evidence is consistent with the component node corresponding to a branch of the candidate chain or is a directly connected component, and the anomaly type combination information in the equipment failure evidence is consistent with or can be covered by the anomaly node combination of that branch, then the equipment failure evidence is considered to be explained by the candidate chain. For example, when the target failure mode of a candidate root cause chain is "bearing wear," and its branches include two paths: "spindle front bearing - increased vibration and impact - bearing wear" and "bearing adjacent area - temperature rise - bearing wear," the equipment failure evidence in step one that belongs to the same analysis window and corresponds to "increased impact on the spindle front bearing accompanied by temperature rise" will be jointly explained by these two paths. However, if there is also equipment failure evidence of "increased current disturbance in the coupling area" in the same window, it will only be included in the explained evidence if there is a support path related to the coupling in the candidate chain. After this processing, a clear one-to-one correspondence is formed between the candidate chain and the evidence. Subsequent calculations no longer stop at the level of "whether the chain is good or bad," but enter the level of "whether the chain truly explains the current anomaly." This is the key difference between the final judgment step of this invention and ordinary sorting output.

[0039] Understandably, the explanatory power of candidate root cause chains for the equipment failure evidence set is quantified using a weighted interpretation consistency model. This model is based on the statistical concept of weighted coverage, which evaluates the degree to which an explanatory object covers the set of observed objects using a weighted summation method. This step builds upon this by extrapolating to the industrial equipment root cause analysis scenario: traditional coverage methods typically treat each covered object with equal weight, but the importance of failure evidence from different equipment in industrial settings varies. The evidence strength calculated in step one... This precisely characterizes the significance of the anomaly; therefore, this step uses the strength of evidence. As a weighting factor, "explaining strong anomalies" is given a higher contribution, while "explaining only weak anomalies" is naturally reduced. Correspondingly, for any candidate root cause chain, the equipment failure evidence within the current analysis window is first traversed line by line. If a piece of evidence is explained by any branch path in that candidate chain, its explanation flag is recorded as 1; otherwise, it is recorded as 0. Then, the evidence is weighted and summed according to its strength, and normalized using the sum of the strengths of all equipment failure evidence within the current window to obtain the explanation consistency score. ; in, This represents the explanatory consistency score of the current candidate root cause chain; This indicates the first candidate root cause chain that is under the same analysis window identifier as the current candidate root cause chain. The strength of evidence for equipment failure directly derives from the standardized equipment failure evidence set output in step one; Indicates the first The explanation flag for each piece of equipment failure evidence relative to the current candidate root cause chain is set to 1 if the equipment failure evidence can be explained by at least one branch path in the current candidate root cause chain, and 0 otherwise; the denominator contains... This represents the sum of the strength of all equipment failure evidence under the same analysis window identifier. The derivation of this formula can be explicitly stated as follows: Step one first provides the strength of each piece of equipment failure evidence. Then, the corresponding evidence is determined through the evidence-by-evidence interpretation relationship in this step. The strengths of all explained equipment failure evidence are summed to form the numerator, which is then normalized to the sum of the strengths of all equipment failure evidence in the current window to obtain the proportion of the current candidate root cause chain that explains the overall anomaly in the window. Taking the results of a centrifugal pump in one analysis window as an example, if step one outputs three pieces of equipment failure evidence, their evidence strengths are respectively... , , Among them, the candidate root cause chain "bearing wear" can explain the first two reasons but not the third, which corresponds to... , , After substituting, we get If the other candidate root cause chain, "coupling misalignment," can only explain the third one, then the corresponding... , , After substituting, we get This result indicates that the "bearing wear" candidate root cause chain explains a significantly higher degree of the major anomalies within the current analysis window. Because... The strength of evidence after normalization in step one. This serves as an interpreter for values ​​of 0 or 1, therefore Since it is a dimensionless ratio, its numerical range naturally falls between 0 and 1.

[0040] Furthermore, in obtaining the explanatory consistency score Then, combine the comprehensive score of the candidate root cause chain obtained in step three. and coverage The final decision score is then formed. This score is based on a multiplicative fusion model. The original idea of ​​multiplicative fusion is that when multiple evaluation dimensions have a "co-established" relationship with the final credibility, using a product form naturally reflects the weakest link effect; that is, a low score in any dimension will significantly lower the overall result. This step follows this idea and specifically extrapolates it to the root cause analysis of industrial equipment failures: Step three... Describe the reliability of candidate root cause chains in terms of structural organization, path quality, and anomaly consistency. Describe the breadth of coverage of the candidate root cause chain to the evidence of current window device failure, and in this step... This describes the explanatory depth of the candidate root cause chain for the main anomalies in the current window; therefore, only when all three are simultaneously high does the candidate root cause chain truly qualify as the final root cause analysis result. Based on this, the final decision score is written as: ; in, This represents the final decision score for the candidate root cause chain; This represents the comprehensive score of the candidate root cause chain obtained in step three; This represents the candidate root cause chain coverage obtained in step three. This represents the consistency score of explanation just calculated in this step. This formula has a clear logical relationship with the previous one: the previous formula first addresses the question of "whether the candidate root cause chain adequately explains the evidence of the current window device failure," thus obtaining... This formula will then be used to... The structured score formed in step three Coverage score The components are combined to obtain the final result used for sorting and judgment. Taking the centrifugal pump mentioned above as an example, if the candidate root cause chain for "bearing wear" has been obtained in step three... , This step calculates the result. Then, after substituting, we have If the candidate root cause chain for "coupling misalignment" corresponds in step three... , This step calculates the result. Then, after substituting, we have Numerical results show that the former is significantly higher than the latter, indicating that "bearing wear" is not only structurally more reasonable and has a wider coverage, but also has a higher explanatory power for the main anomalies in the current window, thus making it more suitable as the final root cause analysis result. If this process is repeated over several historical failure windows, it can usually be observed that the candidate chain corresponding to the true root cause... The values ​​are concentrated in the higher range, while the operating condition disturbances or accompanying phenomena correspond to the candidate chains. The value is significantly low. This phenomenon can be verified by comparing it with the equipment maintenance records, thereby providing a basis for setting the scoring threshold during subsequent deployment.

[0041] After completing the above calculations, according to The candidate chains for industrial equipment failure root causes are sorted from highest to lowest. The failure mode nodes, critical component nodes, and anomalous phenomenon nodes corresponding to the highest-ranked candidate chain are then read to form the industrial equipment failure root cause analysis results. If there is a significant gap between the highest-scoring candidate chain and the second-highest-scoring candidate chain, the failure mode corresponding to the highest-scoring candidate chain is directly identified as the root cause of the target industrial equipment in the current analysis window. Component nodes associated with this candidate chain and repeatedly pointed to by the explained equipment failure evidence are considered as failed components. Specifically, the cumulative strength of the explained equipment failure evidence corresponding to each component node is compared, and the component node with the largest cumulative value is taken as the failed component. If the first two or three candidate chains... If the values ​​are close, they are retained together as parallel candidate results, and then sorted according to... The values ​​indicate priority. The final output includes the fault type, the failed component, and the corresponding explanation chain structure, which can be traced layer by layer back to the path score in step two. and related ratings And further trace back to the strength of evidence of equipment failure in step one. This involves combining information such as component location and anomaly type. The resulting structure is not an isolated classification label, but a complete analytical chain that traces back from the final conclusion to the knowledge graph path and then to on-site anomaly evidence. The final output is the root cause analysis result of industrial equipment failure; when multiple possible causes need to be presented simultaneously, they can also be categorized... The system sorts and outputs several candidate results along with their priorities.

[0042] This invention also provides a knowledge graph-based industrial equipment failure root cause analysis system, the system comprising: The evidence generation module is used to acquire vibration data, temperature data, and operating parameter data of industrial equipment, and synchronously organize them to generate a standardized equipment failure evidence set. Each piece of equipment failure evidence in the standardized equipment failure evidence set includes analysis window identifier, evidence strength, component location, and abnormality type combination information. The evidence generation module is used to map each piece of equipment failure evidence in the standardized equipment failure evidence set to component nodes, abnormal phenomenon nodes, and failure mode nodes based on a pre-built industrial equipment failure knowledge graph. It calculates the association score between a single piece of equipment failure evidence and a candidate failure mode node, and combines the retention score of the corresponding candidate path and the convergence degree of multiple evidences to filter the path and form an evidence association subgraph. The candidate chain generation module is used to merge multiple paths around the same fault mode node to form a root cause candidate chain based on the evidence association subgraph, calculate the comprehensive score and coverage of the root cause candidate chain, the comprehensive score is obtained by combining the mean of the path scores with the structural closure enhancement coefficient and the abnormal conflict suppression coefficient, and the coverage is obtained by the proportion of the number of evidence pieces covered by the candidate chain to the total number of evidence pieces in the window, and generate a root cause candidate set for industrial equipment faults. The root cause output module is used to determine the interpretability score of each root cause candidate chain for the evidence within the window based on the industrial equipment failure root cause candidates and the standardized equipment failure evidence set; combine the comprehensive score, coverage and interpretability score to obtain the final decision score, and output the failure type, failed component and corresponding interpretability chain structure according to the final decision score as the root cause analysis result of industrial equipment failure.

[0043] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0044] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0045] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0046] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A knowledge graph-based method for root cause analysis of industrial equipment failures, characterized in that, The method includes: S1. Acquire vibration data, temperature data, and operating parameter data of industrial equipment, and simultaneously organize them to generate a standardized equipment failure evidence set; wherein, each piece of equipment failure evidence in the standardized equipment failure evidence set includes analysis window identifier, evidence strength, component location, and abnormality type combination information; S2. Based on the pre-built industrial equipment fault knowledge graph, each piece of equipment fault evidence in the standardized equipment fault evidence set is mapped to component nodes, abnormal phenomenon nodes, and fault mode nodes. The association score between a single piece of equipment fault evidence and a candidate fault mode node is calculated. The path is then filtered by combining the retention score of the corresponding candidate path and the convergence degree of multiple evidences to form an evidence association subgraph. S3. Based on the evidence association subgraph, multiple paths are merged around the same fault mode node to form a root cause candidate chain. The comprehensive score and coverage of the root cause candidate chain are calculated. The comprehensive score is obtained by combining the mean of the path scores with the structural closure enhancement coefficient and the abnormal conflict suppression coefficient. The coverage is obtained by the proportion of the number of evidence pieces covered by the candidate chain to the total number of evidence pieces in the window. A root cause candidate set for industrial equipment failure is generated. S4. Based on the industrial equipment failure root cause candidates and the standardized equipment failure evidence set, determine the interpretability score of each root cause candidate chain for the evidence within the window; combine the comprehensive score, coverage and interpretability score to obtain the final decision score, and output the failure type, failed component and corresponding interpretability chain structure according to the final decision score as the result of industrial equipment failure root cause analysis.

2. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 1, characterized in that, The steps for the synchronized organization are as follows: Based on the vibration data, temperature data, and operating parameter data, vibration impact features, temperature change features, and operating condition deviation features are extracted according to the analysis window, and then normalized and weighted fused to generate a standardized equipment fault evidence set. Within each analysis window, the vibration data is used to extract the vibration impact characteristics, specifically the difference between the vibration peak value and the vibration mean value within the current window, which is used to characterize local mechanical impact. Temperature data is used to extract temperature change features, specifically the difference between the current window temperature and the previous window temperature, which is used to characterize short-term temperature rise. The operating parameter data is used to extract the operating parameter offset features, specifically the degree of deviation of the speed, load or current from the center value of the relatively stable operating range within the current window, which is used to characterize the operating condition disturbance level of this window.

3. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 2, characterized in that, The weights for the weighted fusion are preset according to the equipment type and component fault sensitivity; The strength of the evidence is the weighted sum of the vibration and impact characteristics, temperature change characteristics, and operating condition deviation characteristics with their corresponding weights; The standardized equipment failure evidence set includes analysis window identifiers, component locations, anomaly combination semantics, and evidence strength. The anomaly type combination information is the corresponding basic anomaly phenomenon, which is a joint interpretation of vibration and shock, temperature change and operating condition deviation.

4. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 1, characterized in that, The pre-built industrial equipment fault knowledge graph assigns each piece of equipment fault evidence in the standardized equipment fault evidence set to component nodes, anomaly nodes, and fault mode nodes, specifically as follows: For each piece of standardized equipment failure evidence, first locate the corresponding component node in the component index table of the industrial equipment failure knowledge graph based on the component location, and then complete the combination disassembly and abnormal phenomenon node location in the abnormal phenomenon word list based on the abnormal type combination information. Extending from component nodes along component connection relationships to first-order adjacent components, retrieving possible corresponding fault mode nodes from the abnormal phenomenon nodes along fault cause relationships, and generating a candidate node set around the current equipment fault evidence.

5. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 1, characterized in that, The association score is calculated and generated based on the strength of evidence of equipment failure, the anomaly semantic matching coefficient, and the component structural consistency coefficient. Among them, the abnormal semantic matching coefficient is determined by the degree of keyword matching between the abnormal combination in the evidence and the incident abnormal phenomenon node of the candidate fault mode node, and the component structure consistency coefficient is determined by the connection level of the component position in the evidence and the component set associated with the candidate fault mode node in the equipment structure diagram.

6. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 5, characterized in that, The process of combining the retention score of the corresponding candidate path with the multi-evidence convergence degree to filter the path forms an evidence association subgraph, specifically as follows: The retention score is calculated and generated based on the association score of the target failure mode node pointed to by the current candidate path, the evidence convergence coefficient, and the hierarchical length of the current path. The evidence convergence coefficient is the ratio of the number of device failure evidence items that are under the same analysis window identifier as the current candidate path and point to the same target failure mode node to the total number of device failure evidence items under the current analysis window identifier. The level length of the current path is the number of edges traversed from the starting component node to the target failure mode node. Based on preset thresholds and path retention scores, high-scoring nodes and paths are selected to form an evidence association subgraph.

7. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 6, characterized in that, The nodes of the evidence association subgraph include component nodes, anomaly nodes, and fault mode nodes related to the input evidence set, and the edges include component connection relationships, component performance relationships, and fault cause relationships; wherein, each fault mode node is accompanied by association score and path score information contributed by several equipment fault evidence, and each path contains equipment fault evidence.

8. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 1, characterized in that, The comprehensive score is calculated and generated based on the mean retention score of each path in the root cause candidate chain, the structural closure enhancement coefficient, and the abnormal conflict suppression coefficient. The structural closure enhancement coefficient is determined based on the proximity of each component node in the equipment structure diagram, and the abnormal conflict suppression coefficient is determined based on whether the abnormal type combinations of each piece of evidence in the chain support each other. The coverage rate represents the coverage of the root cause candidate chain to the device failure evidence set under the same analysis window identifier. It is calculated based on the number of device failure evidence pieces actually covered by the candidate root cause chain under the current analysis window identifier and the total number of device failure evidence pieces under the same analysis window identifier.

9. The knowledge graph-based root cause analysis method for industrial equipment failures according to claim 1, characterized in that, The interpretation consistency score is calculated based on the sum of the evidence strengths of any device failure evidence under the same analysis window identifier as the current candidate root cause chain and the interpretation identifier of the corresponding device failure evidence relative to the current candidate root cause chain. The sum of the evidence strengths of any device failure evidence under the same analysis window identifier as the current candidate root cause chain is the sum of the evidence strengths of all device failure evidence under the same analysis window identifier; Specifically, when the current equipment failure evidence can be explained by at least one branch path in the current candidate root cause chain, the explanation identifier of the corresponding equipment failure evidence relative to the current candidate root cause chain is set to 1; otherwise, it is set to 0.

10. A knowledge graph-based industrial equipment failure root cause analysis system, characterized in that, The system includes: The evidence generation module is used to acquire vibration data, temperature data, and operating parameter data of industrial equipment, and synchronously organize them to generate a standardized equipment failure evidence set. Each piece of equipment failure evidence in the standardized equipment failure evidence set includes analysis window identifier, evidence strength, component location, and abnormality type combination information. The evidence generation module is used to map each piece of equipment failure evidence in the standardized equipment failure evidence set to component nodes, abnormal phenomenon nodes, and failure mode nodes based on a pre-built industrial equipment failure knowledge graph. It calculates the association score between a single piece of equipment failure evidence and a candidate failure mode node, and combines the retention score of the corresponding candidate path and the convergence degree of multiple evidences to filter the path and form an evidence association subgraph. The candidate chain generation module is used to merge multiple paths around the same fault mode node to form a root cause candidate chain based on the evidence association subgraph, calculate the comprehensive score and coverage of the root cause candidate chain, the comprehensive score is obtained by combining the mean of the path scores with the structural closure enhancement coefficient and the abnormal conflict suppression coefficient, and the coverage is obtained by the proportion of the number of evidence pieces covered by the candidate chain to the total number of evidence pieces in the window, and generate a root cause candidate set for industrial equipment faults. The root cause output module is used to determine the interpretability score of each root cause candidate chain for the evidence within the window based on the industrial equipment failure root cause candidates and the standardized equipment failure evidence set; combine the comprehensive score, coverage and interpretability score to obtain the final decision score, and output the failure type, failed component and corresponding interpretability chain structure according to the final decision score as the root cause analysis result of industrial equipment failure.

Images