An intelligent fmeca risk analysis method for failure propagation link

Abstract

This application discloses an intelligent FMECA risk analysis method for failure propagation paths, belonging to the field of risk assessment technology. The method includes: acquiring the design parameter set and failure criterion set of the target system, and constructing an initial structure of a failure mode knowledge graph based on these; for each node, integrating expert experience knowledge and physical model knowledge to calculate multi-dimensional feature vectors and complete graph construction, significantly improving the analysis confidence in small sample scenarios; constructing a failure propagation network based on the causal relationships between nodes, inputting the failure propagation network and the design parameter set into a dynamic graph neural network for task adaptive updating to generate optimized link weights, achieving rapid dynamic adjustment under different design conditions; finally, calculating the comprehensive failure probability of the failure propagation path based on the optimized link weights and identifying key failure paths, providing clear quantitative decision support for engineering design improvement. This achieves intelligent, dynamic, and accurate assessment of failure risks in complex systems.

Description

Technical Field

[0001] This application relates to the field of risk assessment technology, and in particular to an intelligent FMECA risk analysis method for failure propagation links. Background Technology

[0002] Failure Mode, Effects, and Criticality Analysis (FMECA) is a widely used and important reliability analysis method in the design phase of complex mechanical systems. However, with the increasing functional density and growing complexity of failure mechanisms in modern equipment systems, traditional FMECA methods face challenges in practical engineering applications, including insufficient knowledge integration, inadequate depth of failure correlation analysis, insufficient task adaptability, and limited decision support capabilities. Traditional methods struggle to effectively characterize the multi-level causal coupling relationships between failure modes, cannot dynamically adjust the analysis model according to changes in design parameters, and suffer from fragmented processing stages lacking an integrated feedback mechanism. Therefore, there is an urgent need for an intelligent risk analysis method that can deeply characterize the failure propagation chain and adaptively adapt to different design tasks. Summary of the Invention

[0003] The purpose of this application is to provide an intelligent FMECA risk analysis method for failure propagation links, which can realize dynamic and accurate assessment and decision support for failure risks of complex systems.

[0004] To achieve the above objectives, this application provides the following solution: A smart FMECA risk analysis method for failure propagation chains includes the following steps: Data acquisition steps: Obtain the design parameter set and failure criterion set of the target system.

[0005] The steps for constructing the knowledge graph structure are as follows: Based on the design parameter set and the failure criterion set, an initial structure for the failure mode knowledge graph is constructed; the initial structure includes several nodes, each node corresponding to a failure mode.

[0006] Feature calculation steps: For each node, multi-source knowledge, including at least expert experience knowledge and physical model knowledge, is integrated to calculate the multi-dimensional feature vector of the node, and the multi-dimensional feature vector is filled into the node attributes corresponding to the initial structure to complete the construction of the failure mode knowledge graph.

[0007] Network construction steps: Based on the causal relationships between multiple nodes in the failure mode knowledge graph, construct a failure propagation network containing directed edges and corresponding conditional failure probabilities.

[0008] Dynamic update step: Input the failure propagation network and the design parameter set into the dynamic graph neural network, and use the dynamic graph neural network to perform task adaptive update of the failure propagation network according to the design parameter set to generate optimized link weights.

[0009] Risk analysis steps: Based on the optimized link weights, calculate the comprehensive failure probability of at least one failure propagation path in the failure propagation network, and identify critical failure paths based on the comprehensive failure probability.

[0010] Optionally, the design parameter set includes at least one of functional requirement parameters, structural design parameters, environmental adaptability parameters, and interface constraint parameters; the failure criterion set includes at least one of functional failure criteria, task failure criteria, and interface failure criteria.

[0011] Optionally, based on the design parameter set and the failure criterion set, an initial structure for the failure mode knowledge graph is constructed, including the following steps: The target system is functionally decomposed to obtain multiple functional subsystems.

[0012] For each of the functional subsystems, the potential failure modes of the functional subsystems are enumerated, and each potential failure mode is treated as a node.

[0013] Assign a domain label to each node and establish a mapping relationship between the domain label and the node to obtain the initial structure of the failure mode knowledge graph.

[0014] Optionally, by fusing multi-source knowledge, including at least expert experience knowledge and physical model knowledge, to calculate the multi-dimensional feature vector of the node, the following steps are specifically included: By using the expert experience and knowledge path, the multidimensional feature vector of the node based on expert rating and the corresponding first confidence level are obtained.

[0015] By using the physical model knowledge path, the multidimensional feature vector and the corresponding second confidence level of the node are obtained based on the failure physical model.

[0016] Based on the first confidence level and the second confidence level, the multidimensional feature vectors from the two paths are weighted and fused to obtain the multidimensional feature vector of the node.

[0017] Optionally, when the deviation of the multidimensional feature vectors from the two paths exceeds a predetermined threshold, a verification and correction mechanism is triggered; after the deviation of the multidimensional feature vectors from the two paths has been verified and corrected, the multidimensional feature vectors from the two paths are re-weighted and fused.

[0018] Optionally, the multidimensional feature vector includes at least a severity component to characterize the degree of impact of the failure, an occurrence component to characterize the probability of the failure occurring, and a detection component to characterize the detectability of the failure.

[0019] Optionally, based on the causal relationships between multiple nodes in the failure mode knowledge graph, a failure propagation network containing directed edges and corresponding conditional failure probabilities is constructed, including the following steps: Determine the propagation direction between two nodes with causal relationship in the failure mode knowledge graph to establish the directed edge.

[0020] Based on the occurrence and detection components of the node after failure and the severity component of the node before failure, the conditional failure probability of the directed edge is calculated, and the conditional failure probability is used as the initial weight of the directed edge to obtain the failure propagation network.

[0021] Optionally, the dynamic graph neural network includes a dynamic weight calculation layer and a graph attention convolutional layer; the failure propagation network is adaptively updated based on the design parameter set by the dynamic graph neural network to generate optimized link weights, including the following steps: In the dynamic weight calculation layer, the design parameter set is used as the input condition to calculate the task-related importance weight of each node in the failure propagation network, and then input into the graph attention convolutional layer.

[0022] In the graph attention convolutional layer, for each node, based on the task-related importance weight of the node, the feature information of the node's neighboring nodes is aggregated through the graph attention mechanism to update the node's feature representation.

[0023] Based on the updated feature representations of each node and the task-related importance weights, the weights of each directed edge in the failure propagation network are adaptively adjusted to generate the optimized link weights.

[0024] Optionally, after identifying the critical failure path based on the comprehensive failure probability, the method further includes the following steps: Perturbations are applied to each component of the multidimensional feature vector, and the relative rate of change of the comprehensive failure probability is calculated to perform sensitivity analysis and obtain a list of key influencing parameters.

[0025] Based on the list of key impact parameters, improve design schemes are retrieved and recommended from a pre-set knowledge base of improvement measures, and the risk reduction effect after implementing the improve design schemes is predicted.

[0026] Optionally, the method further includes a closed-loop feedback step: Acquire verification data based on monitoring during testing or actual operation; Using the verification data as a reference, at least one of the following is corrected: the multidimensional feature vector of the node, the parameters of the failure physics model, and the network parameters of the dynamic graph neural network.

[0027] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides an intelligent FMECA risk analysis method for failure propagation chains. First, the design parameter set and failure criterion set of the target system are acquired to lay the data foundation for subsequent analysis. Then, based on the design parameter set and failure criterion set, an initial structure of a failure mode knowledge graph is constructed, enabling systematic organization of failure modes. Next, for each node, multi-source knowledge such as expert experience and physical model knowledge is integrated to calculate its multi-dimensional feature vector and fill it into node attributes to complete the knowledge graph construction, thereby effectively improving the analysis confidence in small sample scenarios. Then, based on the causal relationships between nodes in the knowledge graph, a failure propagation network containing directed edges and corresponding conditional failure probabilities is constructed to deeply characterize the multi-cascade failure propagation process, breaking through the limitations of traditional tabular analysis. Next, the failure propagation network and design parameter set are input into a dynamic graph neural network. This network adaptively updates the failure propagation network according to the design parameter set to generate optimized link weights, thereby achieving rapid dynamic adjustment under different design conditions and significantly reducing the workload of repetitive analysis. Finally, based on the optimized link weights, the comprehensive failure probability of the failure propagation path is calculated and key failure paths are identified, providing clear quantitative decision support for engineering design improvement. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart of an intelligent FMECA risk analysis method for failure propagation links provided in an embodiment of this application.

[0030] Figure 2 This is a flowchart of step S2 in an intelligent FMECA risk analysis method for failure propagation links provided in an embodiment of this application.

[0031] Figure 3 This is a flowchart of step S3 in an intelligent FMECA risk analysis method for failure propagation links provided in an embodiment of this application.

[0032] Figure 4This is a flowchart of step S4 in an intelligent FMECA risk analysis method for failure propagation links provided in an embodiment of this application. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] This application provides an intelligent FMECA risk analysis method for failure propagation chains. In one exemplary embodiment, a separation mechanism in the aerospace field is used as a specific application object for detailed explanation. Figure 1 As shown, the method includes steps S1 to S6. The overall method is divided into six stages, and the main inputs, calculation processes, and output results of each stage will be described in detail below.

[0036] S1. Data acquisition steps: Obtain the design parameter set and failure criterion set of the target system.

[0037] Specifically, extract the complete set of design parameters for the object to be analyzed from the design document. D params These include: functional requirements parameters (such as motor rated voltage 28V, peak torque 4.5N·m, rotation angle 90°), structural design parameters (such as spring preload 180N, spring constant 12N / mm, thread specification M8×1.25, protective coating Ni-P chemical plating thickness 25μm, mating clearance design value 0.05~0.12mm), environmental adaptability parameters (such as operating temperature range -55℃ to +85℃, impact load not exceeding 500g / 11ms), and interface constraint parameters (such as cumulative operating life target 8000 hours).

[0038] At the same time, the set of invalidation criteria is clearly defined. F criteria This refers to the quantitative criteria for determining whether a system can perform its intended function under specified operating conditions. These criteria include: functional failure criteria (mechanism unlocking time exceeding 200ms), task failure criteria (inability to complete separation action), and interface failure criteria (fitting gap exceeding 0.05~0.18mm). These criteria explicitly define the quantitative boundaries between "failure" and "normal," ensuring that the analysis scope only includes failure modes that lead to unacceptable functionality.

[0039] In addition, this step also outputs an analysis requirements specification, such as the analysis level being the subsystem level and the acceptable threshold for the overall failure probability being set to 0.15.

[0040] S2. Graph Structure Construction Steps: Based on the design parameter set and the failure criterion set, construct the initial structure of the failure mode knowledge graph; the initial structure includes several nodes, each node corresponding to a failure mode. For example... Figure 2 As shown, step S2 specifically includes the following sub-steps: S21. Decompose the target system functionally to obtain multiple functional subsystems. Based on the principle of functional independence, decompose the target system functionally into several functional subsystems (such as electric drive subsystem, transmission subsystem, coating and interface subsystem, and overall functional subsystem), and clarify the interface relationships between each subsystem.

[0041] S22. For each functional subsystem, enumerate the potential failure modes of the functional subsystem and treat each potential failure mode as a node. Perform failure mode analysis independently on each functional subsystem, systematically enumerating potential failure modes from multiple dimensions such as material failure, structural failure, functional failure, and interface failure.

[0042] S23. Assign a domain label to each node and establish a mapping relationship between the domain label and the node to obtain the initial structure of the failure mode knowledge graph. Based on the engineering characteristics of each failure mode, classify them into the corresponding engineering domain (such as electric drive system, materials engineering, mechanical engineering, overall design), and establish a mapping relationship between failure mode, engineering domain, and expert domain, which serves as an index for subsequent expert rule conditional triggering.

[0043] Finally, each failure mode is presented as a node. Each node is identified by its failure mode type as its node ID, uses the three-dimensional feature vector calculated in step S3 as its node attribute, its engineering domain as its node type label, and the extracted key parameters as its additional attributes, thus establishing the node layer of the failure mode knowledge graph. It should be noted that the knowledge graph established in step S2 is only an initial skeleton; before step S3 is executed, the three-dimensional feature vector attributes of each node are filled with zero vectors.

[0044] In an exemplary embodiment, engineers divided the separation mechanism in the aerospace field into four subsystems (electric drive subsystem, transmission subsystem, coating and interface subsystem, and overall functional subsystem), identifying a total of 15 failure mode nodes. The node IDs, names, subsystems, and key parameters of six typical nodes are as follows: FM-01 "Inter-turn short circuit in motor windings" (electric drive subsystem, key parameter: winding insulation resistance), FM-02 "Spring preload decay" (transmission subsystem, key parameter: spring constant change rate), FM-03 "Thread surface wear" (transmission subsystem, key parameter: thread contact surface wear), FM-04 "Coating peeling and degradation" (coating and interface subsystem, key parameter: remaining coating thickness), FM-05 "Match clearance out of tolerance" (coating and interface subsystem, key parameter: measured fit clearance value), and FM-06 "Unlocking time out of limit" (overall functional subsystem, key parameter: measured unlocking time). The current SOD (Solution Mode Occurrence) for each node is placed at (0,0,0). The node attribute filling process proceeds in parallel with subsequent step S3.

[0045] S3. Feature Calculation Steps: For each node, multi-source knowledge, including at least expert experience knowledge and physical model knowledge, is integrated to calculate the multi-dimensional feature vector of the node. This multi-dimensional feature vector is then filled into the node attributes corresponding to the initial structure to complete the construction of the failure mode knowledge graph. For example... Figure 3 As shown, this step specifically includes: S31. Obtain the multi-dimensional feature vector of the node based on expert rating and the corresponding first confidence level through the expert experience knowledge path.

[0046] S32. Obtain the multi-dimensional feature vector and corresponding second confidence level of the node calculated based on the failure physical model via the physical model knowledge path.

[0047] S33. Based on the first confidence level and the second confidence level, the multidimensional feature vectors from the two paths are weighted and fused to obtain the multidimensional feature vector of the node. When the deviation of the multidimensional feature vectors from the two paths exceeds a predetermined threshold, a verification and correction mechanism is triggered; after the deviation of the multidimensional feature vectors from the two paths is verified and corrected, the multidimensional feature vectors from the two paths are weighted and fused again.

[0048] This step is the core computational component of this application's scheme. The multidimensional feature vector of a node includes three components: severity S, occurrence O, and detection D. Fusion is achieved through the following two parallel paths: Path 1, Expert Experience and Knowledge Path: First, the expert database is loaded, and the professional domain identifiers, years of experience, and historical accuracy of each expert are read. Initial weights for each expert are calculated according to predetermined rules. For each failure mode, a subset of experts matching the domain is selected from the expert database based on its engineering domain label. Each expert independently assigns an integer score (from 1 to 10) and a confidence level for the failure mode's S, O, and D components according to the scoring rules in the rule base. After collecting the expert scores, a weighted average and a consensus score are calculated. When the consensus score is lower than a preset threshold, a scoring coordination process is initiated, feeding back the aggregation results to experts with significant scoring discrepancies. The aggregation is then iteratively adjusted and repeated until the consensus requirement is met. The output of this path is the expert-weighted three-dimensional score vector for each failure mode. S exp , O exp , D exp ) and the corresponding first confidence level C exp .

[0049] Path Two, Physical Model Knowledge Path: The system initializes a physical failure model library containing quantitative physical failure models for different failure mechanisms. For example, for fatigue failure, based on the material's SN fatigue life curve and Miner's cumulative damage criterion, the system inputs the stress amplitude and number of operating cycles under the current condition to calculate the cumulative damage value and map it to the fatigue failure probability. For wear failure, based on the Archard wear formula, the system inputs the contact load, relative sliding speed, material hardness, and friction coefficient to calculate the wear rate and cumulative wear amount, and uses the probability of exceeding the allowable wear limit as the wear failure probability. For thermal stress failure, based on the material's thermal expansion coefficient, operating temperature range, and structural constraints, the system calculates the stress distribution under thermal cycling loads to obtain the probability of thermally induced crack initiation and propagation. Each physical model incorporates environmental factors to correct the failure probability and dynamically adjusts the model parameters using the maximum likelihood estimation method based on historical validation data. The output of this path is the three-dimensional score vector of the physical model for each failure mode. S phy , O phy , D phy ) and the corresponding second confidence level C phy .

[0050] Dual-path fusion: The outputs of the two paths are weighted and fused according to their respective confidence levels to calculate the final three-dimensional feature vector (S, O, D) for each failure mode node. Specifically, the S, O, and D components are based on ( S exp Cexp + S phy C phy ) / ( C exp + C phy The calculation is performed in a rounded manner. When there is a large deviation between the calculation results of the two paths (such as the difference of a certain component exceeding a predetermined threshold of 2), a verification mechanism is triggered, and relevant experts are notified to re-examine the physical model assumptions or scoring basis. After verification and correction, the results are then fused.

[0051] Taking node FM-04 "coating peeling and degradation" as an example, the fusion process is illustrated as follows: In path one, expert A (15 years of experience, historical accuracy 0.88) scores (S=5, O=8, D=7) with a confidence level of 0.88; expert B (9 years of experience, historical accuracy 0.75) scores (S=5, O=9, D=6) with a confidence level of 0.75; the results are then weighted and aggregated using a weighting of 0.54 / 0.46 to obtain ( S exp =5, O exp =8.4, D exp =6.6), C exp =0.82. In path two, the physical wear model, with a Ni-P coating hardness of 560 HV, contact pressure of 18 MPa, and relative sliding speed of 0.02 m / s, calculates that the cumulative wear over 8000 hours is approximately 22 μm (reaching 88% of the coating thickness of 25 μm), which is mapped to ( S phy =5, O phy =8, D phy =5), C phy =0.79. Two paths O Difference 0.4 D The difference is 1.6, all within the threshold of 2, so they are directly weighted by confidence level and fused (weights 0.51 / 0.49): S =5, O =round(8.4×0.51+8×0.49)=8, D =round(6.6×0.51+5×0.49)=6, therefore the final feature vector of FM-04 is ( S =5, O =8, D =6).

[0052] The same process was used to calculate all 15 nodes, and the typical node results are as follows: FM-01(8,5,4), FM-02(6,7,5), FM-03(7,6,4), FM-04(5,8,6), FM-05(7,5,3), FM-06(9,4,3). These feature vectors were then filled into the attributes of each node in the knowledge graph established in step S2, and all zero-placement vectors were replaced with actual values. At this point, the knowledge graph was fully constructed.

[0053] S4. Network Construction Steps: Based on the causal relationships between multiple nodes in the failure mode knowledge graph, construct a failure propagation network containing directed edges and corresponding conditional failure probabilities. For example... Figure 4 As shown, this step specifically includes: S41. Determine the propagation direction between two nodes with causal relationship in the failure mode knowledge graph to establish the directed edge.

[0054] S42. Calculate the conditional failure probability of the directed edge based on the occurrence degree component and detection degree component of the node after failure and the severity component of the node before failure.

[0055] S43. Using the conditional failure probability as the initial weight of the directed edge, a failure propagation network is obtained.

[0056] In this step, an initial directed adjacency matrix for the failure propagation network is first constructed based on the statistical correlation between nodes of each failure mode. The statistical correlation is derived from the frequency of co-occurrence of two failure modes in historical failure case data, as well as prior information about causal links from engineering experience.

[0057] For each directed edge (node ​​A → node B) in the failure propagation network, calculate the conditional failure probability P(B|A) of node B given that node A has already failed. This conditional failure probability is determined by three factors: the degree component O of node B. B The fundamental failure probability represented (through the mapping function p) O (Convert the O value to a probability value); the detection component D of node B. B The risk of missed detection reflected (the higher the detection rate, the greater the risk of missed detection, as indicated by the coefficient k) D (Magnified); Severity component S of node A A The determined propagation intensity adjustment factor (via coefficient f) S (Magnified). A combined calculation of the three yields P(B|A)=p O (O B )×k D (D B )×f S (S A ), which serves as the edge weight of the directed edge (A→B).

[0058] Then, based on the conditional failure probabilities between the nodes, directed edges are established between all causally related node pairs in the failure mode knowledge graph to form a complete multi-level causal coupling failure propagation network.

[0059] Taking this embodiment as an example, the directed edge FM-04→FM-03: p O =0.40 (O=6 mapping for FM-03), k D =1.05 (D=4 mapping of FM-03), f S =1.00 (S=5 mapping of FM-04), P(FM-03|FM-04)=0.40×1.05×1.00=0.42. Directed edge FM-03→FM-06: p O =0.35 (O=4 for FM-06), k D =1.09 (D=3 for FM-06), f S =1.17 (S=7 for FM-03), P(FM-06|FM-03)=0.35×1.09×1.17=0.45. Directed edge FM-03→FM-05: P(FM-05|FM-03)=0.35; Directed edge FM-05→FM-06: P(FM-06|FM-05)=0.61. Based on these edge weights, multi-cascade failure propagation links can be formed. For example, the initial integrated failure probability of the main path FM-04→FM-03→FM-06 is 0.42×0.45=0.189; the initial integrated failure probability of the parallel path FM-04→FM-03→FM-05→FM-06 is 0.42×0.35×0.61=0.090.

[0060] S5. Dynamic update step: Input the failure propagation network and the design parameter set into the dynamic graph neural network, and use the dynamic graph neural network to perform task adaptive update of the failure propagation network according to the design parameter set to generate optimized link weights.

[0061] The Dynamic Graph Neural Network (DGNN) used in this step includes a dynamic weight calculation layer and a graph attention convolutional layer (GATConv). The specific processing is as follows: First, input preparation: The failure propagation network is represented as a graph structure input. The node feature matrix X consists of the three-dimensional feature vectors (S, O, D) of each node, and the directed edge weight adjacency matrix A consists of the conditional failure probabilities between each node. Simultaneously, the design parameter vector obtained in step S1 is... D params It is provided as a conditional input to the dynamic weight calculation layer.

[0062] Secondly, the dynamic weight calculation layer uses a design parameter vector. D params As conditional inputs, the task-related importance weight vector of each node is dynamically calculated using a parameterized mapping function (such as a trainable neural network or a linear mapping). W node . W node This indicates the relative contribution of each failure mode node to the overall system risk under the current design parameters. For example, in this embodiment, for design parameters... D params =[28V, 4.5N·m, 180N, 25μm, +85℃, 8000h], the calculated W values ​​are: FM-04 = 1.18 (coating degradation intensifies with temperature cycling), FM-03 = 1.15 (thread wear accumulates with operating time), FM-06 = 1.22 (unlocking the limit is a critical node for task completion), and FM-01 = 0.95 (electric drive failure is not sensitive to current parameters).

[0063] Then, the graph attention convolutional layer uses a multi-layered stacked GATConv module to calculate the attention coefficients of each node to its neighboring nodes using an attention mechanism. Based on these attention coefficients, the feature information of the neighboring nodes is aggregated in a weighted manner, and the high-level feature representation of the current node is iteratively updated. This multi-layered stacking can capture information from multi-hop neighbors, allowing the feature representation of each node to be integrated into the global failure propagation context.

[0064] Finally, the link relationship adaptive update layer updates the node feature representations and importance weights accordingly. W node The link connection weights between nodes are adaptively adjusted. Specifically, based on the original conditional failure probability P(B|A) in step S4, the weights are combined with the task-related importance weights of the nodes. W node The weights of each directed edge are dynamically adjusted to obtain the optimized link weight P'(B|A) = min(P(B|A) × W node (1.20) (Taking the upper limit of 1.20 for pruning). For example, P'(FM-03|FM-04)=0.42×1.15=0.48; P'(FM-06|FM-03)=0.45×1.22=0.54 (Taking the upper limit). After this update, the topology of the failure propagation network remains dynamically consistent with the current design parameters.

[0065] S6. Risk Analysis Step: Based on the optimized link weights, calculate the comprehensive failure probability of at least one failure propagation path in the failure propagation network, and identify critical failure paths based on the comprehensive failure probability. This step specifically includes: Comprehensive Failure Probability Calculation: Based on the optimized link weight P' output in step S5, for each failure propagation path in the failure propagation network, the optimal conditional failure probabilities of each level of directed edge are multiplied sequentially according to the propagation order to calculate the comprehensive failure probability of the link endpoint of that propagation path. For example: Main route FM-04→FM-03→FM-06: P path =P'(FM-03|FM-04)×P'(FM-06|FM-03)=0.48×0.54=0.259.

[0066] Parallel path FM-04→FM-03→FM-05→FM-06: P path =P'(FM-03|FM-04)×P'(FM-05|FM-03)×P'(FM-06|FM-05)=0.48×(0.35×1.08)×(0.61×1.05)=0.116.

[0067] Critical Failure Path Identification: Propagation paths with a combined failure probability exceeding a preset risk threshold (0.15 in this embodiment) are identified as critical failure paths. For example, the path FM-04→FM-03→FM-06 has a probability of 0.259 > 0.15, thus it is identified as a critical failure path, and its confidence interval (e.g., [0.198, 0.321]) is output; while the parallel path 0.116 does not exceed the threshold. The risk level and propagation link structure of each critical path are determined by sorting them from high to low combined failure probabilities.

[0068] Sensitivity analysis: A small perturbation (e.g., +5%) is applied to each component of the three-dimensional feature vector of each node in the failure propagation network. By calculating the relative rate of change of the overall failure probability of the critical failure path before and after the perturbation, the sensitivity of each parameter to the overall risk is quantified. For example, after applying a perturbation to the occurrence degree O component (current value 8) of FM-04, P... path The relative sensitivity of FM-03 increased from 0.259 to 0.286, ranking first with a relative sensitivity of 2.08; the severity S of FM-03 ranked second with a relative sensitivity of 1.34. Arranging these parameters from highest to lowest sensitivity yields a priority list of key parameters that have the greatest impact on overall risk.

[0069] Recommended Improvement Measures: For critical impact parameters with high sensitivity rankings and weak points on critical failure paths, design improvement suggestions are proposed based on a pre-built improvement measure knowledge base, focusing on three dimensions: reducing severity, reducing occurrence, and increasing detection. This recommendation process is executed automatically by the software program: the system determines the improvement direction based on the sensitivity analysis results, retrieves a set of alternative measures from the pre-built improvement measure knowledge base that match the current node's failure type, engineering field, and improvement direction, and automatically sorts them by their estimated risk reduction effect, outputting a priority list. For example, for FM-04 (coating peeling degradation) with a high occurrence rate O (current value 8), the system recommends measure A, "upgrade the Ni-P coating to a DLC hard carbon coating," predicting that the O value will decrease from 8 to 5, and recalculating P. path The value decreased to approximately 0.163, representing a 37% reduction in risk; Measure B, "increasing the frequency of coating inspections (from every 500 hours to every 200 hours)," predicted that the D value would decrease from 6 to 4, and the P value would also decrease. path The risk level drops to approximately 0.201, a reduction of 22%. Measure A has higher priority than B and is recommended for implementation first. The engineer will make the final decision based on this, taking into account engineering constraints.

[0070] Closed-loop feedback and iterative optimization: This method also includes a closed-loop feedback step to achieve continuous evolution of the analysis system. When high-fidelity failure verification data is obtained through experimental verification, simulation analysis, or monitoring during actual operation, the feedback optimization module automatically completes the following update operations: Based on actual observed failure mode severity, occurrence, and detection data (S obs O obs D obs Using as a reference, calculate the deviation between the current three-dimensional feature vector of the corresponding node and the actual observed value, and correct the multi-dimensional feature vector of the node based on the deviation.

[0071] The historical accuracy and weight of each expert in the expert database are updated with the actual expert rating accuracy (the system compares each expert's rating vector with the measured vector of the corresponding failure event to calculate the normalized accuracy).

[0072] The model parameters in the physical failure model library are corrected using actual failure propagation observation data (binary statistical results of subsequent failure of node B under the condition of node A failure). For example, the maximum likelihood estimation method is used to correct the propagation-related parameters in the model so that the model output P(B|A) matches the measured propagation frequency.

[0073] The probability of conditional failure between nodes is recalculated using the updated node feature vectors, and the directed edge weights of the failure propagation network are updated accordingly.

[0074] The dynamic graph neural network is fine-tuned using updated training samples to continuously optimize network parameters.

[0075] Through the above iterative feedback, an adaptive optimization mechanism is formed that drives the collaborative updating of the knowledge graph, failure propagation network, and neural network under the design parameters, enabling the analysis system to continuously improve its accuracy and generalization ability as verification data accumulates.

[0076] By implementing steps S1 to S6 above, a high-confidence failure mode feature vector is constructed through multi-source knowledge fusion, the failure propagation process is deeply characterized through multi-level causal coupling links, task-adaptive failure risk prediction is achieved through dynamic graph neural networks, and continuous optimization of the analysis system is achieved through closed-loop feedback mechanisms, thereby providing complete intelligent technical support for the reliability design of complex mechanical systems.

[0077] Compared with the prior art, the present invention has the following beneficial effects: (1) The organic integration of multi-source knowledge significantly improves the confidence of analysis in small sample scenarios. The conditional fusion mechanism of physical failure model, expert profile rule base and historical data knowledge can adaptively allocate fusion weights according to the reliability of each knowledge source, effectively overcoming the problems of strong subjectivity of single expert scoring and low confidence of analysis due to insufficient small sample data. It is especially suitable for engineering scenarios where data is scarce in the early stage of design.

[0078] (2) In-depth characterization of failure propagation, breaking through the limitations of traditional tabular analysis. Through failure mode knowledge graph and multi-level causal coupling links, it can accurately capture the multi-level propagation process of failure in complex systems, cross-domain coupled failure phenomena and multi-level nonlinear interaction relationships, making up for the inherent defects of traditional FMECA that ignores the correlation propagation of failure due to independent item analysis.

[0079] (3) Task dynamic adaptation significantly reduces the workload of repetitive analysis. Dynamic graph neural networks can automatically adjust node weights and link relationships when design parameters change, eliminating the need to re-execute manual full-process analysis, which significantly improves the practicality of the method in multi-scheme comparison and agile design iteration.

[0080] (4) Clearly define quantitative decision support and improve engineering operability. By identifying critical failure paths, conducting multidimensional sensitivity analysis, and prioritizing improvement measures, quantitative basis and operable suggestions are directly provided for engineering design decisions, thereby enhancing the engineering applicability of FMECA analysis results.

[0081] (5) Integrated closed-loop mechanism to support continuous evolution of the analysis system. The integrated design of feedforward and feedback bidirectional data flow enables the knowledge graph, physical model, expert weights and neural network parameters to be continuously optimized as validation data accumulates, giving the analysis system the ability to adapt and evolve, and ensuring the continuous accuracy of FMECA analysis in long-term operating environment.

[0082] Based on the same inventive concept, this application also provides an apparatus for implementing the aforementioned intelligent FMECA risk analysis method for failure propagation links. The solution provided by this apparatus is similar to the implementation described in the above method, and will not be repeated here.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A failure propagation link-oriented intelligent FMECA risk analysis method, characterized in that, include: Data acquisition steps: Obtain the design parameter set and failure criterion set of the target system; The design parameter set includes at least one of functional requirement parameters, structural design parameters, environmental adaptability parameters, and interface constraint parameters; The failure criterion set includes at least one of the following: functional failure criteria, task failure criteria, and interface failure criteria; The steps for constructing the knowledge graph structure are as follows: Based on the design parameter set and the failure criterion set, an initial structure for the failure mode knowledge graph is constructed; the initial structure includes several nodes, each node corresponding to a failure mode. Feature calculation steps: For each node, multi-source knowledge, including at least expert experience knowledge and physical model knowledge, is fused to calculate the multi-dimensional feature vector of the node, and the multi-dimensional feature vector is filled into the node attributes corresponding to the initial structure to complete the construction of the failure mode knowledge graph; Network construction steps: Based on the causal relationships between multiple nodes in the failure mode knowledge graph, construct a failure propagation network containing directed edges and corresponding conditional failure probabilities; Dynamic update step: Input the failure propagation network and the design parameter set into the dynamic graph neural network, and use the dynamic graph neural network to perform task adaptive update of the failure propagation network according to the design parameter set to generate optimized link weights; Risk analysis steps: Based on the optimized link weights, calculate the comprehensive failure probability of at least one failure propagation path in the failure propagation network, and identify critical failure paths based on the comprehensive failure probability.

2. The failure propagation oriented link-based intelligent FMECA risk analysis method according to claim 1, characterized in that, Based on the design parameter set and the failure criterion set, an initial structure for a failure mode knowledge graph is constructed, including: The target system is functionally decomposed to obtain multiple functional subsystems; For each of the functional subsystems, enumerate the potential failure modes of the functional subsystems, and treat each potential failure mode as a node; Assign a domain label to each node and establish a mapping relationship between the domain label and the node to obtain the initial structure of the failure mode knowledge graph.

3. The intelligent FMECA risk analysis method for failure propagation links according to claim 1, characterized in that, By fusing multi-source knowledge, including at least expert experience knowledge and physical model knowledge, the multi-dimensional feature vector of the node is calculated, specifically including: By using the expert experience and knowledge path, the multi-dimensional feature vector of the node based on expert rating and the corresponding first confidence level are obtained; By using the physical model knowledge path, the multidimensional feature vector and the corresponding second confidence level of the node calculated based on the failure physical model are obtained; Based on the first confidence level and the second confidence level, the multidimensional feature vectors from the two paths are weighted and fused to obtain the multidimensional feature vector of the node.

4. The intelligent FMECA risk analysis method for failure propagation links according to claim 3, characterized in that, When the deviation of the multidimensional feature vectors from the two paths exceeds a predetermined threshold, a verification and correction mechanism is triggered; after the deviation of the multidimensional feature vectors from the two paths is verified and corrected, the multidimensional feature vectors from the two paths are re-weighted and fused.

5. The intelligent FMECA risk analysis method for failure propagation links according to any one of claims 1-4, characterized in that, The multidimensional feature vector includes at least a severity component to characterize the degree of impact of the failure, an occurrence component to characterize the probability of the failure occurring, and a detection component to characterize the detectability of the failure.

6. The intelligent FMECA risk analysis method for failure propagation links according to claim 5, characterized in that, Based on the causal relationships between multiple nodes in the aforementioned failure mode knowledge graph, a failure propagation network is constructed, comprising directed edges and corresponding conditional failure probabilities, including: Determine the propagation direction between two nodes with causal relationship in the failure mode knowledge graph to establish the directed edge; Based on the occurrence and detection components of the node after failure and the severity component of the node before failure, the conditional failure probability of the directed edge is calculated, and the conditional failure probability is used as the initial weight of the directed edge to obtain the failure propagation network.

7. The intelligent FMECA risk analysis method for failure propagation links according to claim 1, characterized in that, The dynamic graph neural network includes a dynamic weight calculation layer and a graph attention convolutional layer; The failure propagation network is adaptively updated based on the design parameter set using the dynamic graph neural network to generate optimized link weights, including: In the dynamic weight calculation layer, the design parameter set is used as the condition input to calculate the task-related importance weight of each node in the failure propagation network and input it into the graph attention convolutional layer. In the graph attention convolutional layer, for each node, based on the task-related importance weight of the node, the feature information of the node's neighboring nodes is aggregated through the graph attention mechanism to update the node's feature representation; Based on the updated feature representations of each node and the task-related importance weights, the weights of each directed edge in the failure propagation network are adaptively adjusted to generate the optimized link weights.

8. The intelligent FMECA risk analysis method for failure propagation links according to claim 5, characterized in that, After identifying critical failure paths based on the comprehensive failure probability, the process also includes: Perturbations are applied to each component of the multidimensional feature vector, and the relative rate of change of the comprehensive failure probability is calculated to perform sensitivity analysis and obtain a list of key influencing parameters. Based on the list of key impact parameters, improve design schemes are retrieved and recommended from a pre-set knowledge base of improvement measures, and the risk reduction effect after implementing the improve design schemes is predicted.

9. The intelligent FMECA risk analysis method for failure propagation links according to claim 3, characterized in that, The method also includes a closed-loop feedback step: Acquire verification data based on monitoring during testing or actual operation; Using the verification data as a reference, at least one of the following is corrected: the multidimensional feature vector of the node, the parameters of the failure physics model, and the network parameters of the dynamic graph neural network.

Images