Multimodal knowledge graph enhanced fmea risk dynamic assessment method and system

Abstract

The application discloses a kind of multimodal knowledge graph enhanced FMEA risk dynamic evaluation method and system, for solving the problems of existing FMEA risk parameter static, neglecting failure propagation and lacking unknown risk discovery capability.The method of the application comprises: obtaining multimodal heterogeneous data, mapping process, failure mode and the like into nodes, and constructing a knowledge graph with propagation edges;Based on real-time state, dynamically correct severity, occurrence and detectability, calculate dynamic risk priority number;Extract node topology features, generate structure-enhanced action priority;Identify abnormal structure of the graph, generate potential failure modes through a large language model and feed back the updated graph;Output structured risk control suggestions.The application realizes the transformation from static single-point analysis to dynamic network evaluation, can actively discover unknown risks, and significantly improve the accuracy, systematicness and foresight of risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent manufacturing and quality management technology, specifically involving a multimodal knowledge graph-enhanced FMEA risk dynamic assessment method and system. Background Technology

[0002] Failure Mode and Effects Analysis (FMEA), as a systematic risk management tool, aims to identify potential failure modes in products or processes, analyze their causes and impacts on the system, and develop corresponding prevention and detection measures. This method is widely used in aerospace, automotive manufacturing, medical devices, and general industrial production, and is one of the core means to ensure product quality and operational reliability.

[0003] Traditional FMEA analysis typically uses structured tables to record information and calculates the Risk Priority Number (RPN) by assessing the severity (S), occurrence (O), and detectability (D) of failure modes, thereby ranking risks. In recent years, to overcome the shortcomings of traditional RPN in terms of discrimination and decision-making logic, the Action Priority (AP) method has been gradually introduced to more intuitively guide the development of risk response measures. However, existing FMEA techniques still have the following significant limitations in practical engineering applications.

[0004] First, the static assignment mechanism of risk parameters causes the analysis results to lag significantly behind the actual operating status of the equipment. The S, O, and D parameters in traditional FMEA rely heavily on the subjective experience of experts and historical data during the analysis phase, and once assigned, they remain fixed. During equipment service, its operating status (such as wear and corrosion) changes dynamically with time and operating conditions, and static risk parameters cannot reflect this real-time deterioration, leading to a disconnect between risk assessment results and the actual risk level on site.

[0005] Secondly, the analytical perspective is isolated, neglecting the cascading propagation effect between failure modes. Traditional FMEA typically treats each failure mode as an independent entity for individual evaluation, failing to effectively model the causal relationships and propagation paths between failures in different processes, subsystems, or components. In complex industrial systems, a seemingly minor node failure can propagate and amplify through the system's coupling relationships, ultimately leading to a major accident. Ignoring this cascading effect results in a severe underestimation of the risk of failure modes at critical propagation nodes in the system, thus missing the optimal entry point for risk control.

[0006] Third, the analysis results heavily rely on expert experience and lack the ability to proactively identify unknown risks. The quality of FMEA analysis is largely limited by the breadth of knowledge and depth of experience of the analysis team. On the one hand, historical knowledge within the enterprise (such as past failure cases, maintenance records, and design change orders) has not been fully utilized in a structured manner, forming "information silos." On the other hand, existing methods mainly analyze defined or already occurred failure modes, lacking effective proactive discovery and generation mechanisms for potential and implicit failure modes that are beyond the scope of existing experience, making it difficult to meet the needs of forward-looking risk prediction in complex manufacturing scenarios.

[0007] In recent years, with the development of artificial intelligence technology, some studies have attempted to improve FMEA by introducing technologies such as machine learning, natural language processing, or knowledge graphs. For example, natural language processing is used to extract information from unstructured text, or simple knowledge bases are built to manage FMEA data. However, these solutions mostly remain at the level of data management and information retrieval, failing to fundamentally solve core problems such as the dynamic response of risk parameters, networked propagation modeling of failure modes, and intelligent generation of potential risks. Similarly, although Large Language Models (LLMs) have demonstrated powerful capabilities in semantic understanding and generation, in current FMEA applications, LLMs still lack deep coupling with structured industrial knowledge graphs and dynamic risk assessment mechanisms. The accuracy and reliability of their generated results are difficult to guarantee, making them unsuitable for direct use in high-risk industrial decision-making scenarios.

[0008] In conclusion, there is an urgent need to propose a novel FMEA analysis method that integrates structured relationship modeling, dynamic data analysis, and intelligent generation capabilities to overcome the shortcomings of existing technologies and improve the accuracy, systematicness, and foresight of risk assessment. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide a multimodal knowledge graph-enhanced FMEA risk dynamic assessment method and system, which aims to solve the problems of static risk parameters, neglect of failure propagation effects and lack of ability to discover unknown risks in traditional FMEA. By constructing a dynamic knowledge graph and risk propagation model, combined with a large language model, the invention achieves dynamic, systematic and forward-looking risk assessment.

[0010] To achieve the above objectives, the present invention provides the following technical solution: This invention first proposes a multimodal knowledge graph-enhanced FMEA risk dynamic assessment method, comprising the following steps: Step 1: Acquire data and construct a structured multimodal knowledge graph: Acquire multimodal heterogeneous data, map FMEA core elements to nodes in a knowledge graph, and map the relationships between entities to directed edges, constructing an FMEA risk knowledge graph where nodes carry multimodal fusion features and edges have association propagation weights; wherein, the multimodal heterogeneous data includes at least historical FMEA text, real-time equipment sensor time-series signals, and process structured data; the FMEA core elements include at least process, failure mode, failure cause, failure impact, and control measures; the directed edges include propagation edges used to describe the cascading propagation effect of failure across nodes; Step 2: Dynamically adjust risk parameters based on graph network and real-time state: Based on the real-time operating status of each node in the knowledge graph, the severity S, occurrence O, and detectability D of the risk parameters are dynamically adjusted, and the dynamic risk priority number DRPN of each node is calculated by aggregation; wherein, the adjustment of the severity S is calculated based on the network cascading propagation effect in the knowledge graph. Step 3: Generate enhanced priority determination results from fused graph topological features: Based on the Dynamic Risk Priority Number (DRPN), a preliminary Action Priority (AP) classification is performed. The network topology features of the target failure mode nodes in the knowledge graph are extracted, and structural compensation weights are introduced to adjust the preliminary classification, resulting in a structurally enhanced comprehensive priority. ; Step 4: Drive potential risk mining using a large language model with knowledge graph-enhanced hints: Identify the structural features of the abnormal graph in the knowledge graph, transform them into structured context-enhanced prompt words, and input them into a large language model (LLM) to generate potential undefined failure modes and their associated impacts and control measures. Step 5: Close the loop to update the graph and output the global control strategy: The potential risk content generated by the large language model is injected back into the knowledge graph as new nodes and edges to update it, and the updated global situation and the comprehensive priority are then considered. All failure modes are reordered, and structured dynamic risk control recommendations are output.

[0011] Furthermore, in step one, the initial weights of the propagation edges are calculated based on a weighted fusion of historical conditional probabilities and semantic similarity: in: Represents a node To the node The initial weights of the propagation edges; Nodes in historical data After the node occurs The conditional probability that follows; The cosine similarity of the semantic descriptions of the two nodes; and They are nodes and nodes Semantic description; For weight fusion.

[0012] Furthermore, in step two, a time decay mechanism is introduced when the occurrence degree is 0 in the dynamic correction, incorporating both the current operating status of the equipment and its maintenance history into the calculation: in: This is the dynamic occurrence correction value for node i at time t; The baseline occurrence rate is a static score. The real-time state anomaly mapping function maps the current sensor data deviation to the isolation forest algorithm. Abnormal scores within a given interval; This is a time decay correction factor; This is the timestamp of the last maintenance. The decay time constant is determined based on the MTBF of the device; This refers to the state-risk sensitivity coefficient. This is the attenuation amplitude coefficient; For time.

[0013] Furthermore, in step two, the detection degree D is dynamically corrected based on the confidence model of the detection equipment. : in: This is the dynamic detectivity correction value for node i at time t; The baseline detectivity static score for node i; The overall confidence score of the detection device at node i at time t; The calibration effectiveness score for the testing equipment is linearly mapped based on the calibration expiration time. The current environmental interference factor score is evaluated based on environmental sensor data such as temperature, humidity, and vibration. The historical detection accuracy score; , and These are the weighting coefficients; For time.

[0014] Furthermore, in step two, when dynamically correcting the severity S, the severity increase of the target node is calculated using the propagation edge structure in the knowledge graph: in: This is the dynamic severity correction value for node i; The baseline severity static score for node i; The severity increase of node i due to the network cascading propagation effect; This represents the upper limit of the severity score. For the node Starting from the propagation edge at the maximum propagation depth The set of all downstream nodes reachable within; Represents a node To the node The initial weights of the propagation edges; For nodes in the graph To the node The shortest path hop count; This is the propagation distance attenuation parameter; This is the amplification factor for the cascade effect.

[0015] Furthermore, in step two, the Dynamic Risk Priority Number (DRPN) is: in: This refers to the dynamic risk priority number; This is the dynamic severity correction value for node i; This is the dynamic occurrence correction value for node i at time t; This is the dynamic detection correction value for node i at time t.

[0016] Furthermore, in step three, the network topology features include: weighted out-degree centrality. Betweenness centrality Number of high-risk nodes that can be reached and local clustering coefficient The structural compensation right The normalized network topology features are weighted and fused to obtain the following: in: For min-max normalization; , , and These are the weighting coefficients; The structurally enhanced synthesis priority Based on structurally enhanced comprehensive risk scoring Sure: in: This refers to the dynamic risk priority number; This is the structural compensation amplification factor.

[0017] Furthermore, in step four, the Graph Local Outlier Factor (G-LOF) algorithm is used to identify anomalous node regions where risk parameters have abruptly clustered, and high-risk cascading paths with dynamic severity and propagation edge weights exceeding preset thresholds are identified based on weighted shortest path search; the local subgraph information of the identified anomalous nodes is combined into the structured context-enhanced prompt words according to a preset template. The graph local outlier factor algorithm, G-LOF, quantifies the degree of anomaly by calculating the ratio of the local risk density of each node to the average risk density of its neighborhood. in: For nodes The degree of abnormality; For nodes Locally reachable density, based on nodes With Calculation of distance for dynamic risk parameters between neighboring nodes; For nodes of The set of nearest neighbor nodes.

[0018] Furthermore, in step five, before injecting the content generated by the Large Language Model (LLM) into the knowledge graph, quality filtering is performed. This is done by calculating the cosine similarity between the semantic vector of the generated content and the existing node vectors in the knowledge graph to remove semantic duplicates, and verifying the integrity of the logical chains to ensure a triple structure containing the cause of failure, failure mode, and impact. This is represented as: in: semantic vectors for generating content With existing node vectors in the graph Cosine similarity between them; and These are semantic vectors obtained by encoding the newly generated content and existing node descriptions using a pre-trained language model, respectively. After injecting the triplet structure into the knowledge graph, causal edges and influence edges are established between corresponding nodes; for new propagation paths identified by the Large Language Model (LLM), propagation edges are established, with their initial weights driven by semantic similarity. in: This represents the initial weight of the new propagation edge established by the new propagation path identified by the large language model; For nodes and nodes Cosine similarity in semantic description; and They are nodes and nodes Semantic description; Maintain a dynamic confidence score for each new node generated for each large language model. Its update rules are as follows: in: To observe the relationship between nodes in the actual production data up to time t. The cumulative number of fault events that match the description; This refers to the total number of inspections conducted during the same period. It is a smoothing coefficient used to prevent excessive fluctuations in confidence levels when there is insufficient data in the early stages.

[0019] This invention also proposes a system for implementing the FMEA risk dynamic assessment method with multimodal knowledge graph enhancement as described above, comprising: The multimodal data fusion layer is used to acquire and process historical FMEA text, real-time equipment sensor timing signals, and process structured data; The knowledge graph construction layer is used to map the processed data into nodes and edges, and to construct and store the FMEA risk knowledge graph. The dynamic risk assessment layer is used to dynamically adjust risk parameters based on real-time operating status and calculate the dynamic risk priority number DRPN. The priority evaluation and decision output layer is used to generate the structurally enhanced comprehensive priority. It also outputs structured control recommendations; The potential failure mode identification layer is used to call the large language model, generate potential undefined failure modes and their associated information, and trigger the closed-loop update of the knowledge graph.

[0020] The beneficial effects of this invention are as follows: The multimodal knowledge graph-enhanced FMEA risk dynamic assessment method of this invention has achieved the following technical effects: (1) Dynamic and accurate assessment: In step two, by using real-time data to dynamically correct the three parameters S, O, and D and aggregate them into DRPN, the lag of the traditional static assignment of FMEA is overcome, so that the risk assessment results can respond to changes in equipment status in real time, which significantly improves the accuracy of the assessment. (2) Systemic risk identification: In step one, by constructing a knowledge graph with "propagation edges" and combining it with the node network topology features (such as betweenness centrality) introduced in step three to enhance AP classification, failure modes located at the critical propagation hub of the system can be effectively identified, avoiding the risk being underestimated, and realizing the leap from "single point analysis" to "network propagation analysis"; (3) Proactive risk warning: In step four, by using a large language model to reason about the abnormal structure in the graph, potential and undefined failure modes and control measures are actively generated and fed back to the graph, breaking through the limitations of traditional methods that rely on known experience, and realizing the proactive discovery of unknown risks and knowledge self-evolution. (4) Decision-making feasibility: The final output of structured risk control recommendations (such as emergency action layer and potential risk layer) provides engineers with clear, traceable and operable decision-making basis, forming a closed loop from risk perception to intelligent decision-making. Attached Figure Description

[0021] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart of the FMEA risk dynamic assessment method enhanced by multimodal knowledge graph according to the present invention; Figure 2 This is a schematic diagram of the three-parameter dynamic correction and DRPN aggregation mechanism. Figure 3 This is a comparative diagram of structural compensation mechanisms. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0023] This embodiment aims to propose a multimodal knowledge graph-enhanced FMEA risk dynamic assessment method and system. Specifically, by uniformly modeling and fusing FMEA data and multimodal data from the production process, information such as processes, failure modes, failure causes, failure effects, and control measures are mapped to nodes and relationships in a knowledge graph, thereby constructing a structured network that can describe the causal relationships of failures. Based on this, the association paths between failure modes are analyzed, and a cross-node risk propagation model is constructed, transforming risk assessment from traditional single-point analysis to association analysis based on the overall structure. During the risk assessment process, this embodiment introduces a data-driven dynamic assessment mechanism to adjust severity, probability of occurrence, and detection capability in real time, and calculates a dynamic risk priority number to obtain a baseline risk value that reflects the actual operating status. Furthermore, the Action Priority (AP) criterion is integrated to prioritize failure modes, and the priority results are enhanced by considering their structural position, propagation range, and relationships in the knowledge graph. This ensures that priority assessment not only relies on a single scoring result but also reflects its impact within the overall risk network. Simultaneously, this embodiment introduces a Large Language Model (LLM). After identifying abnormal paths or high-risk structural features, relevant contextual information is input into the LLM, which generates potential failure modes, corresponding impacts, and control measures based on its semantic understanding capabilities. The generated results are then fed back to a knowledge graph for expansion, thereby enabling supplementary identification of undefined risks. In the final output stage, this embodiment integrates dynamic risk assessment results, enhanced priority determination results, and risk propagation path information to rank each failure mode and generate risk control recommendations. This achieves a shift from traditional static scoring methods to multi-dimensional dynamic analysis and intelligent decision support, thereby improving the accuracy, systematicity, and practical application value of FMEA analysis.

[0024] This embodiment employs a collaborative architecture combining a knowledge graph and a large language model to achieve dynamic risk assessment in FMEA. It utilizes a multi-layered data fusion and intelligent reasoning mechanism to address the problems of static data, insufficient correlation, and difficulty in identifying potential failure modes in traditional FMEA analysis methods. The architecture of this embodiment includes a multimodal data fusion layer, a knowledge graph construction layer, a dynamic risk assessment layer, a potential failure mode identification layer, and a priority assessment and decision output layer. Specifically, firstly, through multimodal data acquisition and cleaning, FMEA data and real-time data from the production process are fused and standardized. Next, a knowledge graph is constructed based on the processed data to describe failure modes, failure causes, and their associated impacts. Then, using the knowledge graph and a data-driven dynamic risk assessment mechanism, the risks of each failure mode are corrected and updated in real time. Next, combining the Action Priority (AP) criterion and risk propagation path, the priority of each failure mode is output. Finally, leveraging the generative capabilities of the large language model, undefined failure modes are supplemented, and control measure suggestions are generated to achieve intelligent decision support for the production process.

[0025] Specifically, such as Figure 1 As shown in the figure, the FMEA risk dynamic assessment method enhanced by multimodal knowledge graph in this embodiment includes the following steps.

[0026] Step 1: Acquire data to construct a structured multimodal knowledge graph.

[0027] Acquire multimodal heterogeneous data, map the core elements of FMEA to nodes in a knowledge graph, and map the relationships between entities to directed edges, constructing an FMEA risk knowledge graph where nodes carry multimodal fusion features and edges have association propagation weights; wherein, the multimodal heterogeneous data includes at least historical FMEA text, real-time equipment sensor time-series signals, and process structured data; the core elements of FMEA include at least process, failure mode, failure cause, failure impact, and control measures; the directed edges include propagation edges used to describe the cascading propagation effect of failure across nodes.

[0028] Specifically, in step one, multimodal heterogeneous data such as historical FMEA text, real-time equipment sensor time-series signals, and process structured data are extracted. Each modality of data is independently encoded using a pre-trained language model, a time-series feature encoder, and a graph structure parser to obtain a unified-dimensional multimodal node feature representation. Based on this, processes, failure modes, failure causes, failure effects, and control measures are mapped to knowledge graph nodes. Text semantic features, equipment status features, and process structure features are appended to the corresponding node attributes in the form of multimodal fusion vectors. The causal, impact, control, and cross-process propagation relationships between entities are mapped to four types of directed edges, constructing an FMEA risk knowledge graph where nodes carry multimodal fusion features and edges have propagation weights.

[0029] In the execution of step one, three types of raw data are first collected: historical FMEA text, real-time equipment characteristic data, and process data. Historical FMEA text originates from the enterprise FMEA management platform or PLM system and includes structured and unstructured content such as recorded processes, failure modes, failure causes, failure effects, and control measures from past projects. Real-time equipment characteristic data is continuously collected through industrial IoT interfaces or SCADA systems, including sensor signals such as vibration, temperature, current, and pressure, as well as status characteristics such as equipment runtime and recent maintenance timestamps. Process data describes the technological structure and functional hierarchy of the current product or production line. Next, the five core elements of FMEA are mapped to node types in a knowledge graph, including process nodes, failure mode nodes, failure cause nodes, failure effect nodes, and control measure nodes. Node attributes include the text description of the element, its corresponding process level, and the baseline S, O, D static scores from the historical FMEA. Real-time equipment status characteristics are dynamically attached as attributes to the corresponding process nodes and failure mode nodes, continuously updated with the sampling period.

[0030] The engineering relationships between nodes are mapped to four types of directed edges: causal edges (failure causes pointing to failure modes), impact edges (failure modes pointing to failure effects), control edges (control measures pointing to failure modes or causes), and propagation edges (failure cascade paths across processes). Propagation edges are used to explicitly model the cross-node propagation effect of failure modes at different levels of the system. Their initial weights are calculated based on a weighted fusion of historical conditional probabilities and semantic similarity. in: Represents a node To the node The initial weights of the propagation edges; Nodes in historical data After the node occurs The conditional probability that follows; The cosine similarity of the semantic descriptions of the two nodes; and They are nodes and nodes Semantic description; For weight fusion.

[0031] By introducing propagation edges, the constructed knowledge graph not only stores the local causal relationships of traditional FMEA, but also possesses a topological structure that describes the cascading propagation effect of failures across nodes.

[0032] Step 2: Dynamically adjust risk parameters based on graph network and real-time status.

[0033] Based on the real-time operating status of each node in the knowledge graph, the severity S, occurrence O, and detectability D of the risk parameters are dynamically adjusted, and the dynamic risk priority number DRPN of each node is calculated by aggregation. Among them, the occurrence O is adjusted in combination with the time decay mechanism; the detectability D is updated based on the confidence of the detection device; and the dynamic increase of the severity S is calculated based on the network cascading propagation effect in the knowledge graph.

[0034] Specifically, such as Figure 2 As shown, each failure mode node in the knowledge graph carries a baseline static S, O, D score and real-time device status data. Step two involves dynamically adjusting these three parameters based on the real-time status, and then aggregating and calculating the dynamic risk priority number (DRPN) for each node.

[0035] (1) Occurrence rate O.

[0036] In traditional FMEA, the occurrence degree (O) is statically assigned by experts, failing to reflect the actual degree of equipment degradation. This embodiment introduces a time decay mechanism, incorporating both the current operating status of the equipment and its maintenance history into the dynamic calculation of the O value: in: This is the dynamic occurrence correction value for node i at time t; The baseline occurrence rate is a static score. The real-time state anomaly mapping function maps the current sensor data deviation to the isolation forest algorithm. Abnormal scores within a given interval; This is a time decay correction factor; This is the timestamp of the last maintenance. The decay time constant is determined based on the MTBF of the device; This refers to the state-risk sensitivity coefficient. This is the attenuation amplitude coefficient; For time.

[0037] The core meaning of the above formula is: the longer the equipment has been running since the last maintenance, The larger the value, The corresponding upward movement occurs; the more abnormal the real-time status of the equipment, the higher the risk. The higher the score, Further increase. After the new round of maintenance is completed, Updated to the current time. Resetting it to 1 reflects the proactive risk reduction effect of maintenance actions.

[0038] (2) Detectability D.

[0039] The dynamic correction of detectivity D is based on the confidence model of the detection equipment. Confidence model for testing equipment A comprehensive assessment of the actual reliability of current detection methods: in: This is the dynamic detectivity correction value for node i at time t; The baseline detectivity static score for node i; The overall confidence score of the detection device at node i at time t; The calibration effectiveness score for the testing equipment is linearly mapped based on the calibration expiration time. The current environmental interference factor score is evaluated based on environmental sensor data such as temperature, humidity, and vibration. The historical detection accuracy score; , and These are the weighting coefficients. ; For time.

[0040] The lower the value, the worse the current reliability of the testing equipment. It will rise accordingly.

[0041] (3) Severity S.

[0042] Traditional FMEA's S-value only assesses the impact of a failure mode on its immediate downstream components, neglecting the amplification effect of failure propagation through cascading paths in the system network. If the downstream nodes of the target failure mode node have high severity, high propagation probability, and short propagation paths, their own dynamic severity will increase accordingly. Using the propagation edge structure established in step one, the severity increase of the target node is calculated: in: This is the dynamic severity correction value for node i; The baseline severity static score for node i; The severity increase of node i due to the network cascading propagation effect; This represents the upper limit of the severity score. For the node Starting from the propagation edge at the maximum propagation depth The set of all downstream nodes reachable within; Represents a node To the node The initial weights of the propagation edges; For nodes in the graph To the node The shortest path hop count; This is the propagation distance attenuation parameter; is the amplification factor for the cascade effect, with values ​​ranging from (0, 0.5).

[0043] (4) Dynamic Risk Priority Number (DRPN).

[0044] After dynamically adjusting the three parameters, the dynamic risk priority number is obtained by aggregation: in: This refers to the dynamic risk priority number; This is the dynamic severity correction value for node i; This is the dynamic occurrence correction value for node i at time t; This is the dynamic detection correction value for node i at time t.

[0045] Compared to the traditional static RPN, the dynamic risk priority number DRPN in this embodiment can respond in real time to changes in equipment status, detect equipment degradation and system cascading effects, and provide more accurate basic data for determining the next action priority.

[0046] Step 3: Generate enhanced priority determination results by fusing graph topological features.

[0047] Based on the Dynamic Risk Priority Number (DRPN), a preliminary Action Priority (AP) classification is performed, and the network topology features (such as structural location and propagation range) of the target failure mode nodes in the knowledge graph are extracted. Structural compensation weights are then introduced to adjust the preliminary classification, resulting in a structurally enhanced comprehensive priority. .

[0048] Specifically, such as Figure 3 As shown, the Dynamic Risk Priority Number (DRPN) output in step two simultaneously triggers two parallel calculation paths, which are ultimately merged to generate a structure-enhanced comprehensive priority.

[0049] The first path performs preliminary action priority (AP) classification on all failure mode nodes based on DRPN values. Referring to the AIAG-VDA standard framework, each node is mapped to three preliminary levels: high (H), medium (M), and low (L). Specific thresholds can be configured according to industry standards.

[0050] The second path extracts four network topology features of the target failure mode node from the knowledge graph, capturing the node's structural position within the graph network. Specifically, the network topology features include: weighted out-degree centrality. Betweenness centrality Number of high-risk nodes that can be reached and local clustering coefficient As shown in Table 1.

[0051] Table 1 Definition of Node Network Topology Features The four network topology features are normalized and then weighted and fused to calculate the structure compensation weight. : in: For min-max normalization; , , and Let be the weighting coefficient, satisfying Structural compensation weights Applying this to the DRPN yields a structurally enhanced comprehensive risk score: in: This refers to the dynamic risk priority number; This is the structural compensation amplification factor, with a value of [0,1], used to control the maximum adjustment range of topological features on the final score.

[0052] based on The AP classification is re-executed for all failure mode nodes to obtain the final structural enhancement synthesis priority. Compared to the initial classification based solely on DRPN, It can identify high-structural-risk nodes with moderate DRPN values ​​but located in key propagation hubs in the network, avoiding them from being overlooked due to low initial scores.

[0053] Step 4: Drive potential risk mining using a large language model with knowledge graph-enhanced hints.

[0054] The abnormal graph structure features in the knowledge graph are identified and transformed into structured context-enhanced prompt words, which are then input into a large language model (LLM) to generate potential undefined failure modes and their associated impacts and control measures.

[0055] After completing the comprehensive priority determination in step three, the process enters the anomaly detection phase of this step. The system performs a full scan of the knowledge graph after the dynamic parameter update, and uses the Graph Local Outlier Factor (G-LOF) algorithm to identify anomalous node regions where risk parameters have abruptly clustered. Simultaneously, it identifies high-risk cascading paths where the dynamic severity and propagation edge weights of all nodes exceed preset thresholds based on weighted shortest path search. The G-LOF algorithm quantifies the degree of anomaly by calculating the ratio of the local risk density of each node to the average risk density of its neighborhood. The specific calculation is as follows: in: For nodes The degree of abnormality; For nodes Locally reachable density, based on nodes With Calculation of distance for dynamic risk parameters between neighboring nodes; For nodes of The set of nearest neighbor nodes.

[0056] When the score is significantly greater than 1, it indicates that the risk density around the node is much higher than that of the node itself, representing a local mutation clustering phenomenon. Score exceeds threshold Nodes are marked as mutation cluster nodes, indicating that there may be potential risk patterns around the node that are not yet covered by existing FMEA documents.

[0057] For identified anomalous nodes, the system extracts information from their respective local subgraphs and combines it into structured prompts according to a preset template. These prompts consist of four parts: a description of the current equipment operating conditions, known risk chains in the anomalous subgraph and their dynamic S / O / D parameters, quantitative anomaly indicators such as G-LOF scores and changes in propagation edge weights, and an analysis instruction requiring the model to identify potential undefined failure modes. The structured prompts are then input into the large language model, which is required to output a description of the potential failure mode, the corresponding failure cause, the associated impact path, an initial S / O / D assessment, and control measure recommendations in a structured format.

[0058] Before injecting the content generated by the Large Language Model (LLM) into the knowledge graph, a quality filter is performed. This involves semantic deduplication by calculating the cosine similarity between the semantic vector of the generated content and the existing node vectors in the graph, and verifying the integrity of the logical chains to ensure a triple structure containing the cause of failure, failure mode, and impact. Specifically, the LLM output undergoes two quality filters: first, semantic deduplication is performed by calculating the cosine similarity between the semantic vector of the LLM-generated content and the existing node vectors in the graph. in: semantic vectors for generating content With existing node vectors in the graph Cosine similarity between them; and These are semantic vectors obtained by encoding the newly generated content and existing node descriptions using a pre-trained language model, respectively.

[0059] like Exceeding the threshold The first step is to identify and filter out duplicate content. Next, a logical chain integrity check is performed to ensure that each output contains a complete failure cause-failure mode-impact triplet structure; if any field is missing, a request to generate a replacement is triggered. After these two filtering steps, the output contains verified potential risk content.

[0060] In this embodiment, the output potential risk content is back-injected into the knowledge graph in a standardized triplet format to complete the closed-loop update. Specifically, each potential failure mode is added to the graph as a new node, and the node attributes are assigned values ​​according to the initial S / O / D score output by the LLM, with a confidence label attached. To distinguish existing nodes from those validated by historical data; based on the cause-pattern-influence relationship in the LLM output, causal edges and influence edges are established between corresponding nodes; for new propagation paths identified by LLM, propagation edges are established, with their initial weights set to 0 for historical conditional probability components due to the lack of historical co-occurrence data, and are entirely driven by semantic similarity. in: This represents the initial weight of the new propagation edge established by the new propagation path identified by the large language model; For nodes and nodes Cosine similarity in semantic description; and They are nodes and nodes Semantic description.

[0061] The initial weights will be adjusted and updated according to the complete propagation edge weight formula in step one as real fault data accumulates in subsequent running cycles. The confidence label of the newly injected node will be gradually quantified and improved as real fault data accumulates in subsequent running cycles. The system maintains a dynamic confidence score for each LLM generation node. Its update rules are as follows: in: To observe the relationship between nodes in the actual production data up to time t. The cumulative number of fault events that match the description; This refers to the total number of inspections conducted during the same period. It is a smoothing coefficient used to prevent excessive fluctuations in confidence levels when there is insufficient data in the early stages.

[0062] when Exceeding the preset threshold At that time, the node confidence label was changed by It is updated to Verified_by_Data, and its risk parameters and propagation edge weights are recalculated based on real data and incorporated into a calculation system with the same weight as historical verification nodes.

[0063] Step 5: Close the loop to update the graph and output the global control strategy: The potential risk content generated by the large language model is injected back into the knowledge graph as new nodes and edges to update it, and the updated global situation and the comprehensive priority are then considered. All failure modes are reordered, and structured dynamic risk control recommendations are output.

[0064] After the graph update, the system re-executes steps two and three for all failure mode nodes, performing DRPN calculation and AP classification on both existing and newly injected nodes to reorder all failure modes. The global reordering is triggered by one of the following three conditions: the number of newly injected nodes in a single instance exceeds a preset threshold. The DRPN change of any node exceeds 30%; reaching the system's preset periodic global recalculation cycle. .

[0065] Ultimately, the system outputs structured, dynamic risk control recommendations to FMEA engineers, comprising three levels: the emergency action level lists current... The system provides all failure modes and immediate action recommendations for high-risk nodes; the potential risk layer lists newly identified potential failure modes in this cycle by LLM, sorted by Conf score, and recommends manual confirmation by engineers; the trend warning layer provides early warnings for nodes where DRPN has been rising recently but has not yet reached the high-risk threshold, enabling engineers to intervene in advance.

[0066] After engineers confirm or revise the output suggestions, their operation records (acceptance, rejection, or modification) are captured as high-quality human verification data and fed back to the knowledge graph through the multimodal data access process in step one. This continuously improves the accuracy and coverage of the knowledge graph, forming a two-way enhancement closed loop of "knowledge graph enhancing model reasoning - model feeding back to knowledge graph updates".

[0067] This embodiment also proposes a multimodal knowledge graph-enhanced FMEA risk dynamic assessment system to implement the multimodal knowledge graph-enhanced FMEA risk dynamic assessment method described above. Specifically, the multimodal knowledge graph-enhanced FMEA risk dynamic assessment system of this embodiment includes a multimodal data fusion layer, a knowledge graph construction layer, a dynamic risk assessment layer, a priority assessment and decision output layer, and a potential failure mode identification layer. Specifically, the multimodal data fusion layer is used to acquire and process historical FMEA text, real-time equipment sensor time-series signals, and process structured data; the knowledge graph construction layer is used to map the processed data into nodes and edges, construct and store the FMEA risk knowledge graph; the dynamic risk assessment layer is used to dynamically correct risk parameters according to the real-time operating status and calculate the dynamic risk priority number DRPN; the priority assessment and decision output layer is used to generate the structure-enhanced comprehensive priority. It also outputs structured control suggestions; the potential failure mode identification layer is used to call the large language model to generate potential undefined failure modes and their associated information, and trigger the closed-loop update of the knowledge graph.

[0068] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A multimodal knowledge graph-enhanced FMEA risk dynamic assessment method, characterized in that: Includes the following steps: Step 1: Acquire data and construct a structured multimodal knowledge graph: Acquire multimodal heterogeneous data, map FMEA core elements to nodes in a knowledge graph, and map the relationships between entities to directed edges, constructing an FMEA risk knowledge graph where nodes carry multimodal fusion features and edges have association propagation weights; wherein, the multimodal heterogeneous data includes at least historical FMEA text, real-time equipment sensor time-series signals, and process structured data; the FMEA core elements include at least process, failure mode, failure cause, failure impact, and control measures; the directed edges include propagation edges used to describe the cascading propagation effect of failure across nodes; Step 2: Dynamically adjust risk parameters based on graph network and real-time state: Based on the real-time operating status of each node in the knowledge graph, the severity S, occurrence O, and detectability D of the risk parameters are dynamically adjusted, and the dynamic risk priority number DRPN of each node is calculated by aggregation; wherein, the adjustment of the severity S is calculated based on the network cascading propagation effect in the knowledge graph. Step 3: Generate enhanced priority determination results from fused graph topological features: Based on the Dynamic Risk Priority Number (DRPN), a preliminary Action Priority (AP) classification is performed. The network topology features of the target failure mode nodes in the knowledge graph are extracted, and structural compensation weights are introduced to adjust the preliminary classification, resulting in a structurally enhanced comprehensive priority. ; Step 4: Drive potential risk mining using a large language model with knowledge graph-enhanced hints: Identify the structural features of the abnormal graph in the knowledge graph, transform them into structured context-enhanced prompt words, and input them into a large language model (LLM) to generate potential undefined failure modes and their associated impacts and control measures. Step 5: Close the loop to update the graph and output the global control strategy: The potential risk content generated by the large language model is injected back into the knowledge graph as new nodes and edges to update it, and the updated global situation and the comprehensive priority are then considered. All failure modes are reordered, and structured dynamic risk control recommendations are output.

2. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graphs according to claim 1, characterized in that: In step one, the initial weights of the propagation edges are calculated based on a weighted fusion of historical conditional probabilities and semantic similarity: in: Represents a node To the node The initial weights of the propagation edges; Nodes in historical data After the node occurs The conditional probability that follows; The cosine similarity of the semantic descriptions of the two nodes; and They are nodes and nodes Semantic description; For weight fusion.

3. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graphs according to claim 1, characterized in that: In step two, a time decay mechanism is introduced when the occurrence degree is dynamically corrected to 0, incorporating both the current operating status of the equipment and its maintenance history into the calculation: in: This is the dynamic occurrence correction value for node i at time t; The baseline occurrence rate is a static score. The real-time state anomaly mapping function maps the current sensor data deviation to the isolation forest algorithm. Abnormal scores within a given interval; This is a time decay correction factor; This is the timestamp of the last maintenance. The decay time constant is determined based on the MTBF of the device; This refers to the state-risk sensitivity coefficient. This is the attenuation amplitude coefficient; For time.

4. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graph according to claim 1, characterized in that: In step two, the detection degree D is dynamically corrected based on the confidence model of the detection equipment. : in: This is the dynamic detectivity correction value for node i at time t; The baseline detectivity static score for node i; The overall confidence score of the detection device at node i at time t; The calibration effectiveness score for the testing equipment is linearly mapped based on the calibration expiration time. The current environmental interference factor score is evaluated based on environmental sensor data such as temperature, humidity, and vibration. The historical detection accuracy score; , and These are the weighting coefficients; For time.

5. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graphs according to claim 1, characterized in that: In step two, when dynamically adjusting the severity S, the severity increase of the target node is calculated using the propagation edge structure in the knowledge graph: in: This is the dynamic severity correction value for node i; The baseline severity static score for node i; The severity increase of node i due to the network cascading propagation effect; This represents the upper limit of the severity score. For the node Starting from the propagation edge at the maximum propagation depth The set of all downstream nodes reachable within; Represents a node To the node The initial weights of the propagation edges; For nodes in the graph To the node The shortest path hop count; This is the propagation distance attenuation parameter; This is the amplification factor for the cascade effect.

6. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graphs according to claim 1, characterized in that: In step two, the Dynamic Risk Priority Number (DRPN) is: in: This refers to the dynamic risk priority number; This is the dynamic severity correction value for node i; This is the dynamic occurrence correction value for node i at time t; This is the dynamic detection correction value for node i at time t.

7. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graphs according to claim 1, characterized in that: In step three, the network topology features include: weighted out-degree centrality. Betweenness centrality Number of high-risk nodes that can be reached and local clustering coefficient The structural compensation right The normalized network topology features are weighted and fused to obtain the following: in: For min-max normalization; , , and These are the weighting coefficients; The structurally enhanced synthesis priority Based on structurally enhanced comprehensive risk scoring Sure: in: This refers to the dynamic risk priority number; This is the structural compensation amplification factor.

8. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graph according to claim 1, characterized in that: In step four, the Graph Local Outlier Factor (G-LOF) algorithm is used to identify anomalous node regions where risk parameters have abruptly clustered. Based on weighted shortest path search, high-risk cascading paths with dynamic severity and propagation edge weights exceeding preset thresholds are identified. The local subgraph information of the identified anomalous nodes is combined into the structured context-enhanced prompt words according to a preset template. The graph local outlier factor algorithm, G-LOF, quantifies the degree of anomaly by calculating the ratio of the local risk density of each node to the average risk density of its neighborhood. in: For nodes The degree of abnormality; For nodes Locally reachable density, based on nodes With Calculation of distance for dynamic risk parameters between neighboring nodes; For nodes of The set of nearest neighbor nodes.

9. The FMEA risk dynamic assessment method enhanced with multimodal knowledge graph according to claim 1, characterized in that: In step five, before injecting the content generated by the Large Language Model (LLM) into the knowledge graph, quality filtering is performed. This is done by calculating the cosine similarity between the semantic vector of the generated content and the existing node vectors in the knowledge graph to remove semantic duplicates, and verifying the integrity of the logical chains to ensure a triple structure containing the cause of failure, failure mode, and impact. This is represented as: in: semantic vectors for generating content With existing node vectors in the graph Cosine similarity between them; and These are semantic vectors obtained by encoding the newly generated content and existing node descriptions using a pre-trained language model, respectively. After injecting the triplet structure into the knowledge graph, causal edges and influence edges are established between corresponding nodes; for new propagation paths identified by the Large Language Model (LLM), propagation edges are established, with their initial weights driven by semantic similarity. in: This represents the initial weight of the new propagation edge established by the new propagation path identified by the large language model; For nodes and nodes Cosine similarity in semantic description; and They are nodes and nodes Semantic description; Maintain a dynamic confidence score for each new node generated for each large language model. Its update rules are as follows: in: To observe the relationship between nodes in the actual production data up to time t. The cumulative number of fault events that match the description; This refers to the total number of inspections conducted during the same period. It is a smoothing coefficient used to prevent excessive fluctuations in confidence levels when there is insufficient data in the early stages.

10. A system for implementing the FMEA risk dynamic assessment method with multimodal knowledge graph enhancement as described in any one of claims 1-9, characterized in that: include: The multimodal data fusion layer is used to acquire and process historical FMEA text, real-time equipment sensor timing signals, and process structured data; The knowledge graph construction layer is used to map the processed data into nodes and edges, and to construct and store the FMEA risk knowledge graph. The dynamic risk assessment layer is used to dynamically adjust risk parameters based on real-time operating status and calculate the dynamic risk priority number DRPN. The priority evaluation and decision output layer is used to generate the structurally enhanced comprehensive priority. It also outputs structured control recommendations; The potential failure mode identification layer is used to call the large language model, generate potential undefined failure modes and their associated information, and trigger the closed-loop update of the knowledge graph.

Images