Digital twin diagnostic system and method

Abstract

The application relates to a digital twin diagnosis system and method. The digital twin diagnosis system comprises a multi-source data access unit, a symbolic graph knowledge engine, a graph learning analysis engine, a mechanism data collaborative controller and a twin model; the multi-source data access unit generates a running data sequence; the symbolic graph knowledge engine generates a structure constraint composed of a current effective subgraph G t , a dynamic edge mask matrix M t and an edge symbol matrix S; the graph learning analysis engine outputs a fault original probability and a three-state posterior of each unmeasured node; the mechanism data collaborative controller obtains an effective inferred node state record; the symbolic graph knowledge engine obtains a candidate fault set composed of multiple candidate faults; the mechanism data collaborative controller obtains a candidate fault correction score of all candidate faults and calculates a fusion uncertainty to generate a fusion diagnosis result; and the twin model receives and displays the fusion diagnosis result. The application can improve the reliability of diagnosis results under multi-working-condition and partial-state-unmeasurable conditions.

Description

Technical Field

[0001] This application relates to the field of fault detection technology, specifically to a digital twin diagnostic system and method. Background Technology

[0002] Existing digital twin diagnostic systems include a data acquisition module, a fixed topology graph learning model, a fault classifier, and a digital twin model. The data acquisition module aligns sensor data over time and forms node features; the fixed topology graph learning model performs message propagation based on a preset static adjacency topology; the fault classifier outputs the fault category probability; and the digital twin model updates the fault state accordingly.

[0003] However, these digital twin diagnostic systems typically provide knowledge structures as fixed topologies or fixed relational features to the data model, representing a one-way input from the knowledge side to the data side. When changes in equipment operating conditions cause some fault causal relationships to fail, the fixed graph topology cannot dynamically mask the failed directed edges and cannot reflect the currently effective propagation relationships in a timely manner. Furthermore, the graph structure on which the knowledge side generates candidate fault paths may not be consistent with the structure of message propagation actually executed by the data side, resulting in a disconnect between mechanistic constraints and data reasoning. Summary of the Invention

[0004] This application provides a digital twin diagnostic system and method that can improve the reliability of diagnostic results under multiple operating conditions and conditions where some states are unmeasurable.

[0005] According to a first aspect of the embodiments of this application, a digital twin diagnostic system is provided, comprising: a multi-source data access unit, a symbolic graph knowledge engine, a graph learning analysis engine, a mechanism data collaborative controller, and a twin model; wherein, the multi-source data access unit is configured to generate an operational data sequence based on received sensor measurement data, operation / control data, and operating condition characterization data, and to send the operational data sequence to the symbolic graph knowledge engine and the graph learning analysis engine respectively, wherein the operational data sequence includes at least: current operating condition, operation / control data, valid measured data, measured node status, and unmeasured node identifier; the symbolic graph knowledge engine is configured to generate a current valid subgraph G based on the multi-operating condition symbolic directed graph knowledge model, according to the current operating condition. t Dynamic edge mask matrix M tThe structural constraints, consisting of the edge symbol matrix S, are sent to the mechanism data collaborative controller. The mechanism data collaborative controller then distributes the structural constraints to the graph learning analysis engine. The graph learning analysis engine performs inference based on the running data sequence and the structural constraints, outputting the original failure probability and the three-state posterior of each untested node, and sends the original failure probability and the three-state posterior of the untested node to the mechanism data collaborative controller. The mechanism data collaborative controller calculates the node state uncertainty based on the three-state posterior of the untested nodes, filters out valid inferred node state records based on the node state uncertainty, the unique maximum posterior probability, the confidence threshold, and the uncertainty threshold, and sends the valid inferred node state records back to the symbolic graph knowledge engine. The symbolic graph knowledge engine is further used to analyze the measured node states, the valid inferred node state records, and the current valid subgraph. The system performs evidence weighting, structural reachability judgment, and symbol compatibility reasoning to obtain a candidate fault set consisting of multiple candidate faults, and sends it to the mechanism data collaborative controller. The mechanism data collaborative controller is also used to perform hierarchical correction on the original probability of the fault according to the attribution of each candidate fault in the candidate fault set, obtain the candidate fault correction score of all candidate faults, calculate the fusion uncertainty, generate the fusion diagnosis result, and send the fusion diagnosis result to the twin model. The twin model is used to receive and display the fusion diagnosis result.

[0006] In one implementation, the symbolic graph knowledge engine is further configured to establish a predetermined correspondence between the state variable nodes, operation / control variable nodes, and fault event nodes in the multi-condition symbolic directed graph knowledge model and the corresponding nodes in the graph learning and analysis engine; and to establish message propagation edges in the graph learning and analysis engine based on the predetermined correspondence and the directed edge constraints in the multi-condition symbolic directed graph knowledge model, so that the causal edges in the multi-condition symbolic directed graph knowledge model correspond to the message propagation edges in the graph learning and analysis engine.

[0007] In one implementation, the reasoning based on the running data sequence and the structural constraints includes: In the message generation phase of graph learning message passing, the graph learning analysis engine uses the dynamic edge mask matrix M... t The edge symbol matrix S and graph attention weights constrain message propagation, wherein the dynamic edge mask matrix M... tThe edge symbol matrix S is used to determine whether the corresponding directed edge is allowed to participate in message propagation. The edge symbol matrix S is used to determine the same-direction or opposite-direction feature transformation adopted by the directed edge that is allowed to participate in message propagation. The graph attention weight is used to adjust the propagation strength of messages that are allowed to propagate through the dynamic edge mask matrix and whose feature transformation type has been determined according to the edge symbol matrix. In the l-th level message passing at the current inference time t, the message passed from node i to node j is determined according to the following conditions: in, For the current reasoning moment No. The message passed from node i to node j in the layer. For node i in the th... Layer node characteristics, The dynamic side mask matrix The middle corresponds to the current inference time. Directed edge from node i to node j matrix elements, Based on the edge symbol Choose the same or opposite feature transformation, For directed edges The corresponding edge symbol, For graph attention weights, Determined based on the following conditions: in, For the first The positive edge transformation matrix of the layer, For the first Negative edge transformation matrix of the layer.

[0008] In one implementation, the step of filtering valid inferred node state records based on the node state uncertainty, unique maximum posterior probability, confidence threshold, and uncertainty threshold includes: taking the one with the highest probability among the three posterior probabilities of the untested node as the unique maximum posterior probability; determining that the unique maximum posterior probability is greater than or equal to the confidence threshold, and that the node state uncertainty is less than or equal to the uncertainty threshold; and generating valid inferred node state records.

[0009] In one implementation, the step of performing evidence weighting, structural reachability judgment, and symbolic compatibility reasoning to obtain a candidate fault set consisting of multiple candidate faults includes: performing evidence weighting, structural reachability judgment, and symbolic compatibility reasoning to obtain evidence sufficiency, symbolic compatibility results, and candidate fault scores; and obtaining a candidate fault set consisting of multiple candidate faults based on the evidence sufficiency, the symbolic compatibility results, and the candidate fault scores. Candidate Fault The sign compatibility result is calculated according to the following formula: in, The actual state symbol of the observation node q. Candidate fault f k Predicted state symbol for the node, symbol compatibility results C k It reflects the degree of consistency between the predicted state symbol and the actual state symbol. Candidate Fault Sufficientness of evidence and candidate fault score Calculate according to the following formulas respectively: in, Candidate fault The corresponding set of observation nodes For the set of observation nodes Number of nodes Let q be the evidence weight. The evidence benchmark weights for the measured nodes.

[0010] In one implementation, the candidate fault set includes a priori candidate set, a structurally reachable candidate set, a candidate set with insufficient evidence, and a symbol-incompatible candidate set. The step of performing hierarchical correction on the original fault probabilities is based on the following conditions: in, , For the prior candidate set, For structurally reachable candidate sets, This is a candidate set with insufficient evidence. For the candidate set of symbol incompatibility, The candidate fault correction score is the value for the k-th candidate fault. Let k be the original probability of the k-th candidate fault, where k is a positive integer.

[0011] In one embodiment, calculating the fusion uncertainty includes: calculating the sum of the candidate fault correction scores of all candidate faults; taking the difference between 1 and the sum as the fusion uncertainty; comparing the maximum candidate fault correction score among all candidate fault correction scores with the fusion uncertainty; and if the fusion uncertainty is less than the maximum candidate fault correction score, taking the candidate fault corresponding to the maximum candidate fault correction score as the preferred candidate fault.

[0012] In one implementation, the candidate fault set further includes a symbol-compatible candidate set. The calculation of fusion uncertainty further includes: if the fusion uncertainty is less than the maximum candidate fault correction score, selecting the candidate fault corresponding to the maximum candidate fault correction score as the preferred candidate fault; if the preferred candidate fault belongs to the symbol-compatible candidate set, determining the preferred candidate fault as the current fusion diagnosis fault, and marking its diagnosis status as compatible; if the preferred candidate fault belongs to the insufficient evidence candidate set, marking its diagnosis status as insufficient evidence and pending confirmation; if the preferred candidate fault belongs to the symbol-incompatible candidate set, or if the prior knowledge does not satisfy a branch or structurally unreachable branch, retaining the corresponding knowledge-side attribution identifier and using it as a result to be reviewed. The proportion of all original fault probabilities falling into the symbol-compatible candidate set is defined as a knowledge consistency index. If the knowledge consistency index is less than a knowledge consistency threshold... θ K Furthermore, the maximum original fault probability among all original fault probabilities is higher than the model's high confidence threshold. θ p In the case of low knowledge consistency / high model confidence, an additional low knowledge consistency flag is added, and the diagnostic status is marked as pending review.

[0013] According to a second aspect of the embodiments of this application, a digital twin diagnostic method is provided. The digital twin diagnostic method includes: generating an operational data sequence based on received sensor measurement data, operation / control data, and operating condition characterization data, wherein the operational data sequence includes at least: current operating condition, operation / control data, valid measured data, measured node status, and unmeasured node identifier; and generating a current valid subgraph based on a multi-operating condition symbolic directed graph knowledge model according to the current operating condition. Dynamic edge mask matrix The structural constraints are composed of the edge symbol matrix S; reasoning is performed based on the running data sequence and the structural constraints to output the original failure probability and the three-state posterior of each untested node; the node state uncertainty is calculated based on the three-state posterior of the untested nodes, and valid inferred node state records are obtained by filtering based on the node state uncertainty, the unique maximum posterior probability, the confidence threshold, and the uncertainty threshold; based on the measured node state, the valid inferred node state records, and the current valid subgraph... The system performs evidence weighting, structural reachability judgment, and symbol compatibility reasoning to obtain a candidate fault set consisting of multiple candidate faults; based on the attribution of each candidate fault in the candidate fault set, it performs hierarchical correction on the original probability of the fault to obtain the candidate fault correction score of all candidate faults and calculates the fusion uncertainty to generate a fusion diagnosis result; and receives and displays the fusion diagnosis result.

[0014] In one implementation, the symbolic graph knowledge engine is further configured to establish a predetermined correspondence between the state variable nodes, operation / control variable nodes, and fault event nodes in the multi-condition symbolic directed graph knowledge model and the corresponding nodes in the graph learning and analysis engine; and to establish message propagation edges in the graph learning and analysis engine based on the predetermined correspondence and the directed edge constraints in the multi-condition symbolic directed graph knowledge model, so that the causal edges in the multi-condition symbolic directed graph knowledge model correspond to the message propagation edges in the graph learning and analysis engine.

[0015] The beneficial effects of this application include: 1) By using the same set of currently valid subgraphs, dynamic edge mask matrix and edge symbol matrix to uniformly limit the candidate propagation path on the knowledge side and the message propagation edge on the data side, both sides can carry out reasoning based on the valid structure under the current operating conditions.

[0016] 2) The dynamic edge mask matrix, edge symbol matrix, and attention weights control whether edges propagate, select unidirectional or inverse feature transformations, and adjust message strength, respectively, so that the mechanistic structure constraints are stably effective during the message generation stage, while preserving the model's ability to learn data correlations.

[0017] 3) The graph learning analysis engine simultaneously outputs the original probability of failure and the three-state posterior of untested nodes in the same constrained reasoning; the collaborative controller calculates the uncertainty of node states based on the three-state posterior and only sends back the valid inferred node state records that have passed the confidence and uncertainty gating to the knowledge side.

[0018] 4) Node state uncertainty is used to reduce the weight of presumed evidence and is passed to symbol compatibility, evidence sufficiency and candidate fault score; fusion uncertainty and candidate fault correction score jointly participate in fusion judgment, thereby distinguishing the diagnostic results of specific fault candidate dominance and uncertainty dominance, and retaining the knowledge side attribution basis of the candidate.

[0019] 5) Through the twin state synchronization interface, the current operating conditions, current effective subgraphs, node states and their uncertainties, fused fault scores, fused uncertainties, fused diagnosed faults or preferred candidate faults and their diagnostic status are synchronized to the twin model, so as to achieve a unified representation of operating status, fault candidates and diagnostic results. Attached Figure Description

[0020] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, in which: Figure 1 This is a schematic diagram illustrating the composition of a digital twin diagnostic system according to an exemplary embodiment.

[0021] Figure 2 This is a schematic diagram illustrating the generation, bilateral use, and distribution process of a dynamic edge mask matrix, edge symbol matrix, and currently valid subgraph under a current operating condition, according to an exemplary embodiment.

[0022] Figure 3 This is a schematic diagram of graph learning message passing under the constraints of a dynamic edge mask matrix and an edge symbol matrix, according to an exemplary embodiment.

[0023] Figure 4 This is a schematic diagram illustrating an effective process for generating and transmitting estimated node status records, correcting fault scores, calculating fusion uncertainty, and making fusion diagnostic decisions, according to an exemplary embodiment.

[0024] Figure 5 This is a symbolic directed graph illustrated according to an exemplary embodiment.

[0025] Figure 6 This is a block diagram illustrating an electronic device according to an exemplary embodiment. Detailed Implementation

[0026] It should be understood that in the embodiments of this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0027] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0028] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0029] One embodiment of this application provides a digital twin diagnostic system that enables operational data, mechanistic constraints, inferences about the state of untested nodes, and their uncertainties to participate in diagnosis through a unified data flow.

[0030] Please see Figure 1 The digital twin diagnostic system may include: a multi-source data access unit, a symbolic graph knowledge engine, a graph learning analysis engine, a mechanistic data collaborative controller, and a twin model. In some embodiments, the digital twin diagnostic system may also include an online update unit. For ease of description, the "mechanistic data collaborative controller" will be referred to as the "collaborative controller"; the symbolic graph knowledge engine and its executed mechanistic knowledge reasoning process will be referred to as the "knowledge side," and the graph learning analysis engine and its executed graph learning reasoning process will be referred to as the "data side."

[0031] refer to Figure 1The multi-source data access unit receives sensor measurement data, operation / control data, and operating condition characterization data. After time alignment, measurement validity checks, current operating condition determination, and node mapping, it forms an operating data sequence, which is then provided to the symbolic graph knowledge engine and the graph learning analysis engine. The operating data sequence includes at least the current operating condition, operation / control data, valid measured data, measured node states, and unmeasured node identifiers. The symbolic graph knowledge engine generates structural constraints based on the current operating condition, consisting of the current valid subgraph, a dynamic edge mask matrix, and an edge symbol matrix. This constraint is then sent to the graph learning analysis engine via the mechanism data collaborative controller. The graph learning analysis engine performs constrained inference based on the operating data sequence and structural constraints, outputting the original fault probability and the three-state posterior of unmeasured nodes. The collaborative controller calculates the node state uncertainty based on the three-state posterior and filters valid inferred node state records, sending them back to the symbolic graph knowledge engine. The symbolic graph knowledge engine performs symbolic compatibility re-inference based on the measured node states and valid inferred node state records, and returns the candidate fault set and compatibility results to the collaborative controller. The collaborative controller corrects the fault scores based on the knowledge-side re-inference results and calculates the fusion uncertainty. Subsequently, the fusion uncertainty is compared with the candidate fault correction scores, and the fusion diagnosis result is formed by combining the candidate fault set attribution, and then synchronized to the twin model.

[0032] In the above process, the same currently valid subgraph, dynamic edge mask matrix, and edge symbol matrix simultaneously constrain both knowledge-side candidate propagation and data-side message propagation. Valid inferred node state records, after confidence and uncertainty gating, serve only as supplementary evidence for the knowledge side and are not written back to the graph learning input for the current period. The knowledge-side re-inference results are only used to correct the original fault probabilities obtained in the current period, and the fused uncertainty is used to characterize the degree of diagnostic uncertainty for faults that, after correction, have not yet been clearly attributed to a specific candidate fault.

[0033] like Figure 1 As shown, the multi-source data access unit forms a unified operational data sequence; the symbolic graph knowledge engine generates structural constraints and completes symbolic compatibility re-inference; the graph learning analysis engine outputs the original probability of the fault and the three-state posterior of the untested node under the structural constraints; the collaborative controller completes the calculation of node state uncertainty, the screening of effective inferred node state records, the correction of fault scores, the calculation of fusion uncertainty, and the fusion diagnosis decision; the twin model receives the final fusion diagnosis result.

[0034] ① This indicates that the collaborative controller sends structural constraints to the graph learning analysis engine. These structural constraints are generated by the symbolic graph knowledge engine and sent to the mechanistic data collaborative controller, which then sends them to the graph learning analysis engine. ② This indicates that the graph learning analysis engine returns the original probabilities and the three-state posteriors of untested nodes to the collaborative controller. ③a This indicates that the collaborative controller sends valid inferred node state records back to the symbolic graph knowledge engine. ③b This indicates that the symbolic graph knowledge engine returns the symbolic compatibility re-inference results to the collaborative controller. ④ This indicates that the fused fault scores, fused uncertainties, diagnostic results, and node state information are synchronized to the twin model. These logical interfaces together constitute the bidirectional collaborative system connection between the symbolic graph knowledge engine (knowledge side) and the graph learning analysis engine (data side).

[0035] In some embodiments, the multi-source data access unit is used to generate an operation data sequence based on the received sensor measurement data, operation / control data, and operating condition characterization data, and send the operation data sequence to the symbolic graph knowledge engine and the graph learning analysis engine respectively. The operation data sequence includes at least one of the following: current operating condition, operation / control data, valid measured data, measured node status, and unmeasured node identifier.

[0036] In some embodiments, the multi-source data access unit receives sensor measurement data, operation / control data, and operating condition characterization data, aligns them according to a unified time identifier, and checks the measurement validity. Then, based on the operating condition characterization data and / or operation / control data, it determines the current operating condition, maps the valid measured data (i.e., the portion of sensor measurement data that has been aligned to a unified time identifier and checked for measurement validity) to state variable nodes, and forms measured node states for use by the knowledge side according to preset benchmarks and thresholds. For state variable nodes where valid measured data cannot be obtained at present, an unmeasured node identifier is generated. After processing, an operating data sequence for the current inference cycle is formed. The operating data sequence includes at least the current operating condition, operation / control data, valid measured data, measured node states, and unmeasured node identifiers. This operating data sequence is simultaneously sent to the symbolic graph knowledge engine and the graph learning analysis engine. The symbolic graph knowledge engine uses the current operating condition and measured node states, while the graph learning analysis engine uses the operation / control data, valid measured data, and unmeasured node identifiers to construct node features.

[0037] For example, the measured node status can be obtained by calculating the standardized residual from valid measured data and comparing it with a preset dead zone threshold. When the standardized residual is greater than the positive dead zone threshold, the measured node status is +1; when the standardized residual is less than the negative dead zone threshold, the measured node status is -1; when the residual is between the positive and negative dead zone thresholds, the measured node status is 0. Nodes with invalid sensor data verification do not generate a measured node status and are treated as unmeasured nodes.

[0038] In some embodiments, the symbolic graph knowledge engine receives the current operating condition from the running data sequence, calls a pre-stored multi-condition symbolic directed graph knowledge model, determines the directed edges that can participate in propagation in the current cycle based on the operating condition enabling conditions of each directed edge, and reads the qualitative influence symbols of the corresponding edges. The processing result forms structural constraints, including the currently valid subgraph G. t Dynamic edge mask matrix M t The symbolic matrix S is used to construct the structure constraints, which are sent to the graph learning analysis engine via the collaborative controller and are also used for candidate propagation path selection on the knowledge side. The nodes of the multi-condition symbolic directed graph knowledge model include equipment state variable nodes, operation / control variable nodes, and fault event nodes, and the directed edges represent the causal relationships between the nodes.

[0039] Current valid subgraph G t At the current inference moment Based on the current operating conditions The dynamic edge mask matrix M in the original multi-condition symbolic directed graph knowledge model is transformed. t The remaining subgraph after removing the masked directed edges. Dynamic edge mask matrix M. t It is the current inference moment. A binary matrix used to indicate whether each directed edge is allowed to participate in message propagation / path filtering. Edge symbol matrix. It stores each directed edge The matrix of qualitative causal influence symbols is a matrix predefined within the multi-condition symbolic directed graph knowledge model and does not change with real-time conditions.

[0040] refer to Figure 2 At the current inference moment The symbolic graph knowledge engine, based on the current operating conditions... Determine the directed edge one by one. Enable conditions Determine if the condition is met, and generate a dynamic edge mask matrix. : when At times, there is a side Effective and allowed to propagate under current operating conditions; when At times, there is a side It is invalid and is blocked during this cycle. This corresponds to the directed edge from node i to node j. Applying the method to a multi-condition symbolic directed graph knowledge model, we obtain the current valid subgraph. .

[0041] Each directed edge is pre-configured with operating condition enable conditions. Operating conditions enabling conditions The Boolean discriminant function, determined offline based on the equipment mechanism, is used to determine whether the corresponding causal directed edge is allowed to propagate under a given operating condition; during online inference, the current operating condition is used. Substitute the enabling conditions corresponding to each directed edge. The judgment result is obtained, and a dynamic edge mask matrix is ​​generated. Operating conditions enabling conditions It is not obtained through model training and does not change automatically during online inference.

[0042] edge symbol matrix Used to indicate the qualitative influence direction of a directed edge. This indicates that changes in the source node have the same effect on the target node. This indicates a reverse effect. The edge symbol is predetermined by the engineering mechanism and is not treated as a learnable parameter during training and inference. Only updates are made when the operating conditions change. and its corresponding The qualitative influence sign of the same directed edge remains unchanged.

[0043] The symbolic graph knowledge engine is also used to establish predetermined correspondences between state variable nodes, operation / control variable nodes, and fault event nodes in the multi-condition symbolic directed graph knowledge model and their corresponding nodes in the graph learning and analysis engine. Based on these predetermined correspondences and the directed edge constraints in the multi-condition symbolic directed graph knowledge model, message propagation edges are established in the graph learning and analysis engine, ensuring that causal edges in the multi-condition symbolic directed graph knowledge model correspond to message propagation edges in the graph learning and analysis engine. The same dynamic edge mask matrix is ​​used. It is used on both the knowledge side and the data side: the knowledge side uses it to limit candidate propagation paths, and the data side uses it to limit message propagation edges; the edge symbol matrix S is used for path symbol calculation on the knowledge side and positive and negative edge feature transformation selection on the data side, so that the knowledge side and the data side share the effective edge constraints and edge symbol relationships under the current operating conditions, and respectively complete symbol path reasoning and graph learning message propagation.

[0044] The structural constraints generated by the symbolic graph knowledge engine are used simultaneously by the symbolic graph knowledge engine and the graph learning analysis engine within the current inference cycle; wherein, the dynamic edge mask matrix M t In the symbolic graph knowledge engine, it is used to limit candidate propagation paths, and in the graph learning analysis engine, it is used to limit the message propagation edges that are allowed to participate in message propagation; the edge symbol matrix S is used for symbol calculation of candidate propagation paths in the symbolic graph knowledge engine, and in the graph learning analysis engine, it is used to determine the same-direction or opposite-direction feature transformations corresponding to message propagation edges, so that candidate propagation in the symbolic graph knowledge engine and message propagation in the graph learning analysis engine are based on the same currently valid edge constraints and edge symbol relationships.

[0045] Knowledge side When generating candidate propagation paths, the paths are limited to simple paths without duplicate nodes, and a maximum path length is set to avoid non-convergence of path enumeration in directed graphs containing cycles. Candidate Faults To the same observation node When there is more than one path, first calculate the product of the symbols on each path. If the symbols obtained from all paths are consistent, then use that symbol as the predicted state symbol. If the symbols are inconsistent, the pre-determined exclusionary or optimal processing method used during system deployment will be adopted. The exclusionary processing method marks the node as an inherently conflicting node on the path and excludes it from the observation node set. The optimal processing method selects the preferred path based on the confidence level of the path edge and the path length.

[0046] In some embodiments, reference Figure 3 The graph learning analysis engine receives the running data sequence formed by the multi-source data access unit, as well as the structural constraints (currently valid subgraphs) issued by the collaborative controller. Dynamic edge mask matrix The graph learning analysis engine first constructs initial node features based on the running data sequence, which serve as input to the graph neural network (GNN) (i.e., graph learning network) used internally by the engine. Then, it performs constrained message propagation only along the currently enabled edges. For state variable nodes with valid measured data, features are constructed using one or more of the following: current measurement value, relative baseline deviation, rate of change, frequency domain features, or time-frequency features. For running / control variable nodes, features are constructed using one or more of the following: control commands, setpoints, execution feedback, and their changes. For state variable nodes with unmeasured node identifiers, the measured features are set to zero and the unmeasured node identifier is written simultaneously. For fault event nodes, fault type embedding is used. After inference, the graph learning analysis engine outputs the original fault probability and the three-state posterior of each unmeasured node and sends them to the cooperative controller.

[0047] The graph neural network takes initial node features and structural constraints as input, performs constrained message propagation inference, and outputs the original fault probability and the three-state posterior of each unmeasured node. A dynamic edge mask matrix is ​​used to constrain the available directed edges (i.e., causal edges) for message propagation, and an edge symbol matrix is ​​used to introduce mechanistic constraints on the direction of causal influence during message transmission. The original fault probability is the probability of fault occurrence corresponding to each fault event node, and the three-state posterior of unmeasured nodes is the posterior probability distribution of each unmeasured equipment state variable node on {+1,0,-1}. The three states include positive deviation (+1), normal (0), and negative deviation (-1). For equipment state nodes without sensor measurements, the graph neural network calculates the probability of occurrence for each of the three states: positive deviation, normal, and negative deviation. This set of probabilities is called the three-state posterior.

[0048] In the message generation phase of graph learning message passing, the graph learning analysis engine uses the dynamic edge mask matrix M... t The edge symbol matrix S and graph attention weights constrain message propagation, where the dynamic edge mask matrix M... t The edge symbol matrix S is used to determine whether the corresponding directed edge is allowed to participate in message propagation. The same-direction or opposite-direction feature transformation adopted by the directed edge that is allowed to participate in message propagation is used. The graph attention weight is used to adjust the propagation strength of messages that are allowed to propagate through the dynamic edge mask matrix and whose feature transformation type has been determined according to the edge symbol matrix.

[0049] In the In layer message passing, the message sent by source node i to target node j is simultaneously constrained by the dynamic edge mask matrix, edge symbols, and attention weights: In the formula, For the current reasoning moment No. The message passed from node i to node j in the layer. For node i in the th... Layer node characteristics; Control whether the message is propagated; (.) represents the edge symbol. Choose the same or opposite feature transformation, For directed edges Corresponding edge symbols; is the graph attention weight, and is a non-negative attention weight, used only to adjust the message strength of the enabling edge.

[0050] The feature transformation corresponding to the edge symbol is selected according to the following formula: In the formula, and The first The layer contains positive and negative edge transformation matrices. The dynamic edge mask matrix, edge symbols, and attention weights control propagation permission, transformation type, and message strength, respectively. Attention weights are normalized only within the effective incoming neighbors of the target node and cannot be adjusted. The re-entry of directed edges into propagation cannot change The determined feature transformation type.

[0051] The positive and negative edge transformation matrices can be configured using either strong or soft constraints. Under strong constraints, Directly constraining the two to satisfy an inverse relationship; under the soft constraint method, , Learn separately and set the reverse consistency regularity term L. oppThis is used to limit the degree to which the two deviate from their opposite relationship: This regularization term only constrains the parameter relationships during the message generation stage and does not require that the input and output features of each sample be strictly reversed in numerical value.

[0052] In some embodiments, the graph learning analysis engine employs both fault category supervision and node state symbol supervision during training. Fault category supervision uses fault category labels confirmed through experiments, field tests, maintenance records, or other reliable methods. Node state symbol supervision uses a partial node supervision approach, calculating the corresponding supervision loss only for nodes with valid measured node states, experimental test states, or states confirmed through maintenance records; nodes lacking reliable state labels are not calculated for node state supervision loss. Unconfirmed model-presumed node states are not used as supervision labels. The dynamic edge mask matrix and edge symbol matrix are generated by the symbolic graph knowledge engine and serve as fixed constraint inputs, not participating in the learning process.

[0053] After message passing is completed, the graph learning analysis engine maps the obtained node features through the fault classification output layer and the node state output layer, respectively, and then normalizes them to obtain two types of results: one is the fault set. The raw probability of all fault categories Another category is the posterior distribution of the signs of the unmeasured nodes in the three states of positive deviation, normal, and negative deviation: (4) Among them, This represents the output of the current inference cycle graph learning and analysis engine. This represents the set of untested nodes.

[0054] For each untested node, the graph learning analysis engine determines the state to be inferred and its confidence level based on the three-state posterior, and sends the complete three-state posterior of the untested node to the collaborative controller. The collaborative controller calculates the node state uncertainty according to equation (5). In equation (5), U q The uncertainty of the state of the untested node q. The more concentrated the posterior probability of a certain state, the greater the uncertainty of U. q The closer U is to 0, the closer the posterior probabilities of the three states are. q The closer it is to 1, the more effective it becomes. This uncertainty is used simultaneously for filtering node state records and reducing the weight of knowledge-side evidence.

[0055] In some embodiments, the collaborative controller receives the raw fault probability and the three-state posterior of untested nodes output by the graph learning analysis engine. The collaborative controller calculates the node state uncertainty based on the three-state posterior, and determines whether the node can form supplementary evidence based on the unique maximum posterior probability, confidence threshold, and uncertainty threshold. If the conditions are met, a valid inferred node state record is generated. The valid inferred node state record is sent back to the symbolic graph knowledge engine; nodes that do not meet the conditions remain in the untested state. The raw fault probability is not sent back to the graph learning analysis engine but is retained by the collaborative controller for subsequent fault score correction.

[0056] Specifically, when a node has a unique maximum posterior state and the corresponding confidence level is not lower than the confidence threshold. Furthermore, the uncertainty of the node state does not exceed the preset uncertainty threshold. When the node is in a state of uncertainty, a valid inferred node state record is generated. This means taking the highest probability among the three posterior states of the untested node as the unique maximum posterior probability; ensuring that the unique maximum posterior probability is greater than or equal to the confidence threshold, and that the node state uncertainty is less than or equal to the uncertainty threshold; and then generating a valid inferred node state record.

[0057] A valid estimated node state record should include at least the state symbol, estimated identifier, confidence level, node state uncertainty, and time-to-live (TTL). i and generate operating condition identifiers The result is sent back to the symbol graph knowledge engine; otherwise, the node remains untested and the uncertainty information of the node's state is recorded.

[0058] Among them, the confidence threshold is used to filter out results with low maximum a posteriori probabilities, the uncertainty threshold is used to exclude results with overly dispersed three-state posterior distributions, and the unique maximum a posteriori probability condition is used to eliminate parallel ambiguities. The validity period is used to limit the retention period of valid presumed node state records, and the generated working condition identifier is used to determine whether the working condition structure on which the record is based is still valid.

[0059] In some embodiments, valid presumed node status records and measured node status records are stored separately, without overwriting measured data. When both valid measured node status records and valid presumed node status records exist for the same node, the measured node status is used for both symbol compatibility re-inference and evidence weight calculation. A valid presumed node status record will be invalid if it exceeds its validity period, or if its generated operating condition identifier is inconsistent with the current operating condition. If there is an inconsistency, the valid presumed node status record becomes invalid; if there is no valid measured node status for the corresponding node, the node is treated as an unmeasured node.

[0060] In some embodiments, after a change in operating condition, the dynamic edge mask matrix and the current valid subgraph may change. The state posterior generated based on the structure corresponding to the operating condition before the switch is no longer used for symbol compatibility judgment under the new operating condition. In subsequent inference cycles, the corresponding nodes will be determined based on the structure corresponding to the new operating condition. and the established Re-output the posterior state symbol.

[0061] In some embodiments, valid inferred node state records are only sent back to the symbolic graph knowledge engine as supplementary evidence for re-executing symbolic compatibility re-inference; they are not written into the current graph learning input that generates the posterior, nor are they triggered for the current period's graph learning analysis engine to infer again. That is, valid inferred node state records are not added to the initial node features of the graph learning network in the current inference period; and the graph learning analysis engine is not called again to perform inference calculations in the current period. The knowledge-side re-inference results are only used to correct the original fault probabilities already obtained in the current period, thereby avoiding the same inference result repeatedly supporting the model's original judgment within the same inference period.

[0062] In some embodiments, the symbolic graph knowledge engine receives, in the second phase, the measured node states from the runtime data sequence, the valid inferred node state records returned by the cooperative controller, and the currently valid subgraph G. t The process involves weighting evidence, determining structural reachability, and re-executing symbolic compatibility inference. Since the measured node states are derived from sensor data and represent reliable, real-world observational evidence, they receive a higher baseline weight. Valid inferred node state records, calculated by the graph learning analysis engine, serve as supplementary evidence; therefore, the evidence corresponding to these records receives a lower baseline weight and is further reduced based on node state uncertainty. This process generates a candidate fault set, evidence sufficiency, symbolic compatibility results, and candidate fault scores. The re-inference results are then returned to the collaborative controller for subsequent fault score correction.

[0063] In some embodiments, the candidate fault set may include a priori candidate set, a structurally reachable candidate set, a candidate set with insufficient evidence, a symbolically compatible candidate set, and a symbolically incompatible candidate set. The knowledge side performs priori filtering by reading the historical prior occurrence probabilities of various fault types, setting a priori probability threshold, and including faults with a priori probability greater than or equal to the priori probability threshold into the priori candidate set. Faults with a prior probability lower than the aforementioned prior probability threshold are not included in the prior candidate set. The candidate faults that do not meet the prior attribution criteria are retained for subsequent correction of the original fault probability classification. The remaining candidate faults form the prior candidate set. The knowledge side can perform structural reachability based on the currently valid subgraph. Traverse the prior candidate set For each fault event node, check whether there exists a complete and connected causal propagation path from the fault node to each observation node (i.e., all causal edges on the path are valid edges in this period, mask). If a complete propagation path exists, the candidate fault is included in the structural reachability candidate set. The sufficiency of evidence assessment will continue; if no complete propagation path exists, the candidate fault will not be included in the structurally reachable candidate set. The unreachable structural identifier is retained for subsequent hierarchical correction, and the remaining candidate faults form a candidate set of structures that can be reached. Observation nodes are equipment state variable nodes that can provide observational evidence. These include measured observation nodes whose measured node states are obtained based on valid sensor data, and inferred observation nodes whose valid inferred node state records are obtained through screening by the mechanism data-driven controller. The knowledge side can weight the evidence and apply it to the structurally reachable candidate set. For each fault identified, all observed evidence (measured node status and valid presumed node status records) is weighted. Based on the number and weight of evidence, the sufficiency of evidence is calculated and compared with a preset threshold. If the sufficiency of evidence is greater than or equal to the threshold, the corresponding candidate fault is added to the sufficiency of evidence candidate set, and symbol compatibility is checked. If the sufficiency of evidence is less than the threshold, the corresponding candidate fault is added to the insufficiency of evidence candidate set. For candidate sets with insufficient evidence For candidate faults within the given set, no sign compatibility rate is calculated, no sign compatibility or sign incompatibility conclusions are made, and subsequent sign compatibility assessments for that candidate fault are discontinued; the candidate fault is retained in the insufficient evidence candidate set. The system provides subsequent hierarchical correction. The knowledge side can perform symbol compatibility checks; for faults within the candidate set with sufficient evidence, the symbol compatibility rate is only used for those belonging to... And does not belong to Candidates are judged for compatibility; candidates without a complete propagation path cannot have their path symbol product calculated. Candidate faults in the list of those with insufficient evidence are not considered compatible or conflicting based on their sign compatibility rates. For candidate faults that are structurally reachable and have sufficient evidence but do not meet the sign compatibility condition, they are classified into the sign incompatible candidate set. .

[0064] In some embodiments, for candidate faults Observation node set Includes only the following nodes: in the currently active subgraph Reachable; exists that can determine the symbol of the predicted state. Candidate propagation paths must be available; nodes must have measured node states or valid inferred node state records. Nodes without candidate propagation paths or those determined to have inherent path conflicts are not included. Validally estimated node state records are entered. Before that, its validity period and generated operating condition identifier should also be verified.

[0065] The node evidence weight is calculated according to formula (6b). The more dispersed the three-state posterior corresponding to the effective inferred node state record, the better U... q The larger the value, the lower its evidentiary weight; when the uncertainty of the presumed result reaches its maximum, the record does not constitute valid knowledge-based evidence. Candidate Fault The symbol compatibility ratio (i.e., the symbol compatibility result) is: In the formula, The actual state symbol of the observation node q. Candidate fault f k The predicted state symbol for that node. C k Reflects the degree of consistency between predicted and actual symbols; node state uncertainty is addressed through... w q Enter the calculation of compatible evidence weights to avoid the high uncertainty presumed state and the measured node state having the same effect.

[0066] Sufficientness of evidence and candidate fault score Calculate according to equations (6d) and (6e) respectively, where = .because U q Approved w q Enter and The calculation, It also reflects the uncertainty of symbol compatibility, source of evidence, and presumed state; The lower the value, the higher the diagnostic uncertainty of the candidate. in, Candidate fault The corresponding set of observation nodes For the set of observation nodes Number of nodes Let q be the evidence weight. The evidence benchmark weights for the measured nodes.

[0067] The symbolic graph knowledge engine performs structural reachability checks, filtering from all nodes in the graph to generate candidate faults. The corresponding set of observation nodes .

[0068] when When it is empty or the sum of the evidence weights is zero, , and If none are defined, candidates are treated as having insufficient evidence. When the number of observation nodes is less than... The sum of the weights of evidence is lower than Or the number of measured nodes is less than At that time, candidate faults are included No conclusions are drawn regarding symbol compatibility or symbol incompatibility.

[0069] in, The threshold for the number of observation nodes, This is the threshold for the sum of evidence weights. This is the threshold for the number of nodes measured. As the compatibility threshold, The threshold for scoring candidate faults.

[0070] If the set of observation nodes corresponding to a candidate fault is empty, or the number of observation nodes in the set of observation nodes is less than the threshold for the number of observation nodes, or the sum of the evidence weights of the observation nodes is less than the threshold for the sum of the evidence weights, or the number of measured nodes in the set of observation nodes is less than the threshold for the number of measured nodes, then the candidate fault is classified into the insufficient evidence candidate set, and no qualitative judgment of sign compatibility or sign incompatibility is made for the candidate fault.

[0071] For candidate faults that are not included in the insufficient evidence candidate set, a sign compatibility determination is performed. When the sign compatibility rate of the candidate fault is greater than or equal to the compatibility rate threshold, the candidate fault score is greater than or equal to the candidate fault score threshold, and the number of measured nodes in the observation node set is greater than or equal to the number of measured nodes threshold, the candidate fault is assigned to the sign compatible candidate set.

[0072] For candidate faults that do not belong to the insufficient evidence candidate set or meet the admission criteria of the symbol-compatible candidate set, they are classified into the symbol-incompatible candidate set.

[0073] The candidate fault set, consisting of all candidate faults from the insufficient evidence candidate set, the symbol compatible candidate set, and the symbol incompatible candidate set, is output to the mechanism data collaborative controller.

[0074] The candidate fault set satisfies: in, and Non-overlapping. Enter The candidate should not belong to and simultaneously satisfy , The number of measured nodes is not less than The lower weight of presumed evidence, the threshold for the number of measured nodes, and the threshold for candidate fault scoring collectively limit the effect of the presumed state, preventing candidates supported entirely by the presumed state from entering the system. .

[0075] In some embodiments, reference Figure 4 The mechanism-data collaborative controller receives the candidate fault set, evidence sufficiency, and symbol compatibility results returned by the symbolic graph knowledge engine, and then bases the analysis on the candidate faults in... , , , , The attribution of faults in the graph learning inference output is used to perform hierarchical correction on the original fault probabilities. The five branches correspond to prior non-satisfaction, structural inaccessibility, insufficient evidence, sign incompatibility, and sign compatibility, respectively, and the above set relationship ensures that each candidate fault enters only one branch in a correction.

[0076] To reflect the consistency between the output of the graph learning analysis engine and the symbolic compatibility judgment of the knowledge side, a knowledge consistency index K is defined. K represents the percentage of candidates falling into the symbolic compatibility candidate set among all original fault probabilities. The probability percentage: Where K is larger, it indicates a higher probability that the graph learning output obtains knowledge-side symbolic compatibility support; a smaller K indicates a lower probability. Candidates with insufficient evidence cannot be determined as compatible or incompatible, and therefore are not included in the numerator. When the denominator is zero, the knowledge consistency index is undefined, and the system outputs an invalid model probability flag.

[0077] Graded correction is performed according to the following formula: in, Candidates with incompatible symbols are given the heaviest weight reduction, followed by those with unreachable structures, while candidates outside the prior candidate space are given the lightest weight reduction. Candidates with insufficient evidence retain their original probability because the existing evidence is neither sufficient to confirm nor disprove the candidate. Candidates with incompatible symbols are removed during the process. It retains and demotes in weight, while also attaching an incompatible flag. For candidate fault correction scores, Let be the original probability of the k-th candidate fault.

[0078] After graded correction, the sum of the correction scores for each candidate fault may be less than 1. The difference between this sum and 1 is defined as the fusion uncertainty U. F . Among them, U F Let U be the fusion uncertainty. When the original fault probability is normalized and 0 ≤ λ3 < λ2 < λ1 < 1, the correction scores for each candidate fault and the fusion uncertainty U are... F The sum is 1. Fault support reduced by knowledge-side hierarchical correction due to prior dissatisfaction, structural inaccessibility, or sign incompatibility is no longer assigned to other candidates but is retained as fusion uncertainty. U F The larger the value, the lower the support level retained on the specific fault candidate after knowledge-based correction, and the higher the uncertainty of the fusion diagnosis.

[0079] If all candidate fault correction scores are zero, then U F =1, outputting no valid candidate identifier. Otherwise, the collaborative controller will fuse the uncertainty U. F Compare with the maximum correction score among all candidate faults; when U F When the value of UF is not lower than the maximum correction score, it indicates that the current uncertainty is dominant, no preferred candidate fault is formed in this cycle, and a pending confirmation status is output; when UF is lower than the maximum correction score, the candidate fault with the highest correction score is determined as the preferred candidate fault.

[0080] If the preferred candidate fault belongs to the symbol incompatible candidate set, or does not belong to the prior candidate set, or belongs to the prior candidate set but not to the structurally reachable candidate set, the preferred candidate fault will be used as the result to be reviewed.

[0081] If the preferred candidate fault belongs to the insufficient evidence candidate set, the diagnostic status is marked as insufficient evidence, pending confirmation.

[0082] If the preferred candidate fault belongs to the symbol-incompatible candidate set, or does not belong to the prior candidate set, or belongs to the prior candidate set but not to the structurally reachable candidate set, the preferred candidate fault will be used as the result to be reviewed.

[0083] When the preferred candidate fault belongs to the symbol-compatible candidate set F C When this occurs, it is identified as a current fusion diagnostic failure; it belongs to the insufficient evidence candidate set F. U When the evidence is insufficient, it is retained as the preferred candidate fault and marked as insufficient evidence or pending confirmation; when it belongs to the symbol incompatible candidate set, the prior does not satisfy the branch or the structure is unreachable branch, the corresponding knowledge side attribution identifier is retained and it is treated as a result to be reviewed.

[0084] When the knowledge consistency index K < θ K And the maximum original probability of failure is higher than θ. p At the same time, add low knowledge consistency / high model confidence markers and record relevant node and directed edge information for subsequent review.

[0085] The fusion diagnosis results are sent to the twin model via the twin state synchronization interface, including the fusion fault score, fusion uncertainty, current fusion diagnosed fault or preferred candidate fault and its diagnosis status.

[0086] In some embodiments, the online update unit is an optional module used to receive confirmed operational data sequences, maintenance feedback, or labeled samples; update the learnable parameters of the graph learning analysis engine, as well as the prior probabilities of fault nodes and the confidence of directed edges in the symbolic graph knowledge engine; and output updated parameters after offline verification or manual review for use in subsequent inference cycles, without immediately replacing the parameters currently in use within the current inference cycle. The online update unit does not directly modify the edge symbol matrix S. Knowledge adjustments involving causal relationships, edge symbols, or operating condition enabling conditions should be updated according to the knowledge model maintenance process after manual review based on confirmed data and maintenance feedback.

[0087] Example 1: This embodiment uses the operational setting variables and sensor measurement data from the CMAPSS FD002 aero-engine degradation data as online diagnostic inputs, and selects the 69th cycle of test set Unit 4 as the current diagnostic time, i.e., t=69. The current cycle and the previous 29 cycles constitute the X40:69 operating window, and sequentially complete the following processes: operating condition identification, structural constraint generation, constrained graph learning and reasoning, uncertainty gating, knowledge-side re-reasoning, fault score correction, and twin state synchronization.

[0088] The implementation process includes: (1) The multi-source data access unit receives the operation setting variables and sensor measurement data of the 69th cycle and the previous 29 cycles of the FD002 test set Unit 4, forming an X40:69 operation window, and determines the current operation condition as c2 based on the operation setting variables of the 69th cycle. After node mapping, standardization and status determination, the measured node status, unmeasured node identifier and operation data sequence are output.

[0089] (2) The symbolic graph knowledge engine generates the current effective subgraph G by taking the current working condition c2 and the preset working condition enabling conditions as input. 69 Dynamic edge mask matrix M 69 The edge symbol matrix S is used simultaneously for knowledge-side candidate path filtering and data-side graph learning message propagation.

[0090] (3) The graph learning analysis engine constructs the initial features h of the nodes based on the running data sequence. 69 (0), and call the graph learning model trained based on synthetic supervised samples, with {h 69 (0),G 69 M 69Let S be the input for constrained message propagation, and let S be the output for the raw failure probability and the three-state posterior of untested nodes. The model achieves a failure classification accuracy of 83.33% and a node three-state accuracy of 91.67% on the synthetic supervised sample retained validation set.

[0091] (4) The collaborative controller receives the three-state posterior of the untested node, calculates the node state uncertainty, and filters the valid inferred node state records by combining whether the maximum posterior state is unique, the confidence level corresponding to the maximum posterior state, the confidence threshold, and the uncertainty threshold. The inferred results obtained through gating are only sent back to the symbolic graph knowledge engine and are not written back to the periodic graph learning input.

[0092] (5) The symbolic graph knowledge engine integrates the measured node status and the effective inferred node status records, and sequentially completes the structural reachability, evidence sufficiency and symbol compatibility judgments on G69. It outputs the candidate fault set, evidence sufficiency and compatibility results and candidate fault scores, and feeds them back to the collaborative controller.

[0093] (6) The collaborative controller calculates knowledge consistency based on the attribution of candidate fault sets and performs hierarchical correction on the original fault probabilities. This yields p′(fc) = 0.04988 and p′(ft) = 0.75060, with a fusion uncertainty U. F =0.19952. Because f t The candidate fault correction score is higher than U F , will f t As the current preferred candidate fault; f t ∈F U Therefore, the diagnostic status of "insufficient evidence / pending confirmation" is output and synchronized to the twin model.

[0094] (a) Formation of running data sequence The FD002 training and test sets are organized by engine number. With a seed of 42, the training data is divided into a training side and a validation side based on engine number. 208 engines are used for model fitting and parameter calculation, while 52 engines are used for independent verification. The training side is used to fit the operating condition identification model, calculate near-health benchmarks and state determination thresholds, and construct graphical inputs; the validation side is used to verify the above processing results.

[0095] For the t-th operating cycle, the multi-source data access unit reads three operating setting variables and the currently available sensor measurements to form the cycle input vector: X t =[u1,u2,u3,x1,x2,...,xn] Among them, u1, u2, and u3 are operation / control data and operating condition characterization data, which are used to identify the current operating condition and construct operation / control node characteristics; xi is sensor measurement data.

[0096] The multi-source data access unit uses the [u1, u2, u3] of the training engine to fit the StandardScaler and KMeans models, with the number of clusters set to 6, and numbered sequentially from c1 to c6 according to the cluster center values ​​of the first running settings variables; the verification and testing sides directly call the fitted model to determine the working condition of each cycle. The six working condition centers are shown in Table 1.

[0097] Table 1 Online diagnosis employs a causal time window of length 30. For the current period t, the multi-source data access unit extracts only data from the current period and the previous 29 periods, forming: Wt = {Xt-29, Xt-28, ..., Xt} Therefore, the input window corresponding to cycle 69 of test set Unit 4 is X40:69. The runtime setting variables read in cycle 69 are (10.0045, 0.2500, 100.0), which are identified as c2 after condition identification. The final output of this stage is the X40:69 runtime data window and the condition identifier c2, which serve as inputs for subsequent node state construction and dynamic structural constraint generation.

[0098] (ii) Generation of structural constraints The symbolic graph knowledge engine receives the current working condition c2 and calls the pre-established basic symbolic directed graph G=(V,E). This embodiment sets up 12 nodes, such as... Figure 5 As shown: fault event nodes fc, ft; HPC health status node vc; gas path coupling status node vb; turbine response status node vt; measurable state variable nodes P30, T30, T50, EPR; and operating / control variable nodes u1, u2, u3. Among these, P30, T30, T50, and EPR map to FD002 channels s7, s3, s4, and s10, respectively.

[0099] The symbolic graph knowledge engine determines eight basic directed edges and their symbols based on the equipment mechanism: fc → vc = -1, ft → vt = -1, vc → P30 = +1, vc → T30 = -1, vc → vb = +1, vb → vt = +1, vt → T50 = -1, and the cross-layer edge vc → vt = +1. The edge symbols are fixed mechanism constraints and do not participate in learning; whether each edge participates in propagation under the current operating condition is determined by the operating condition enabling condition Γ. (i,j) (c) Decision.

[0100] According to the preset working condition-directed edge enabling rules, the symbolic graph knowledge engine filters basic edges based on working condition c2. As an exemplary configuration, this embodiment stipulates that the cross-layer edge vc→vt is not enabled under working condition c2, i.e., Γvc,vt(c2)=0, while vc→vb and vb→vt remain enabled under working condition c2. Thus, after being identified as working condition c2 in the 69th cycle of Unit 4, according to equation (1), M 69 (vc,vt)=0.

[0101] After the above-described enabling process, 7 valid edges are retained under c2: fc→vc(-1), ft→vt(-1), vc→P30(+1), vc→T30(-1), vc→vb(+1), vb→vt(+1), and vt→T50(-1), generating the current G. 69 and M 69 .

[0102] In G 69 The candidate propagation path adopts a simple path without duplicate nodes, with a maximum number of edges of 6. If there are multiple valid paths from the same candidate fault to the same evidence node, the product of the path symbols is calculated separately. When the symbols are consistent, a definite prediction state is formed. When there is a symbol conflict that cannot be selected, a definite prediction state is not generated.

[0103] At this point, {G} is generated from the input condition c2 in this stage. 69 M 69 The three structural constraints are: S, ..., S. These constraints are directly used for candidate propagation path selection on the knowledge side and then distributed to the graph learning analysis engine via the collaborative controller to limit the propagation of graph learning messages on the data side. The knowledge side limits candidate propagation paths, and the data side limits graph learning message propagation. This results in a unified dynamic structural constraint that can be generated under the same operating conditions and used by both sides for subsequent collaborative inference.

[0104] (III) Construction of Node Features Among them, in obtaining X 40:69 After running the window and the current operating condition c2, the multi-source data access unit standardizes the effective sensor measurements of each cycle within the window according to the operating condition, and calculates the standardized residuals with the near-health data of the training side as a reference, thereby forming the measured node status of the knowledge side and the node characteristics of the data side.

[0105] Using the first 20 operating cycles of each engine on the training side as a near-health reference sample, the near-health mean values ​​of each observation sensor were calculated according to the operating conditions. μ ( c,s ) and standard deviation σ( c,s), and calculate the standardized residuals of each periodic measurement accordingly. When σ( c,s )≤ ε σ At that time, take ε σ =1×10 -12 The sensor was deemed to lack an effective fluctuation scale under the corresponding operating conditions, and the period measurement was marked as invalid; subsequent window mean, trend and threshold statistics only used the effective standardized residual.

[0106] Further, the average value of the standardized residuals for the first 30 cycles of each engine on the training side was calculated, and the 95th percentile of its absolute value was taken as the dead zone threshold for the corresponding observation node. δ s The current window's average standardized residual is greater than... δ s When the node state is less than -1, the node state is recorded as +1; when it is less than -1, the node state is δ s Time thresholds are recorded as -1; all others are recorded as 0. The thresholds for the main observation nodes are shown in Table 2.

[0107] Table 2 The training data was statistically analyzed using the aforementioned standardization and state determination rules. During the degradation process, the standardized residuals of P30 generally showed a negative shift, while T30 and T50 generally showed a positive shift, corresponding to P30↓, T30↑, and T50↑. This result is consistent with the degradation propagation direction described by the preset edge symbols, and the same trend is also observed in the engine-level reserved verification samples, thus completing the independent verification of state determination and symbol relationships.

[0108] During online diagnosis, the multi-source data access unit uses a window with a length of 30. W t Standardized residuals are calculated periodically, generating statistical characteristics such as window mean, trend, window standard deviation, and effective period ratio. Historical window data is then used to construct the characteristics of the current node, while the current operating condition c2 is used to determine the propagation structure for this period.

[0109] The near-healthy reference standard deviation of EPR is 0 or within the rounding range under c1, c2, and c4, and is also relatively small under other operating conditions. To avoid distortion caused by excessively small standard deviations, this embodiment retains EPR plot nodes only when σ(c,s) > ε. σ The periodic calculation standardization value is not used as valid state evidence from the knowledge side. Unit 4 X 40:69 Within the window, EPR satisfies the above validity conditions for 17 out of 30 cycles.

[0110] For Unit 4 X40:69 After the above processing is performed in the running window, the main observation node characteristics are obtained as shown in Table 3.

[0111] Table 3 Therefore, the measured node states P30=0, T30=0, and T50=+1 are formed in this cycle; vc, vb, vt No direct measurements were taken; unmeasured nodes were marked. The standardized residuals for the current period of P30, T30, and T50 are -0.78197, 0.75137, and 3.11932, respectively. u1, u2, u3 The current values ​​after StandardScaler transformation on the training side are -0.94902, -1.03918, and 0.41878, respectively.

[0112] The graph learning analysis engine further encodes the above-mentioned measured states, window statistical features, runtime setting variables, and node type information into 12-dimensional node initial features. h i (0) Among them, dimensions 1 to 6 of the measured state nodes such as P30, T30, T50, and EPR are written to the current... z Value, window mean z, trend, window standard deviation, effective period ratio, and measurement validity indicator; u1, u2, u3 The 7th and 8th dimensions contain the standardized current value and window change amount, and the 12th dimension contains the running / control node identifier; fc, ft The 9th and 10th dimensions are written with fault type embeddings; vc, vb, vt The measured values ​​are set to zero, and the unmeasured node identifier is written in the 11th dimension.

[0113] Finally, the 12 node features are stacked in a fixed order to form a... h 69 (0) and with the currently valid subgraph G 69 Dynamic edge mask matrix M 69 edge symbol matrix S Together with the measurement validity identifier, they serve as inputs to the graph learning analysis engine, which then proceeds to subsequent constrained graph learning inference.

[0114] (iv) Learning and reasoning with constrained graphs After completing the node feature construction, the graph learning analysis engine receives h 69 (0) Current G 69 M 69The edge symbol matrix S is used for inference with a two-layer constrained GNN. The model propagates messages only along the currently enabled edges according to equations (2) and (3a). S selects positive and negative feature transformations and adjusts the strength of effective edge messages through attention weights. The fault classification output head and the node state output head output the original fault probability and the three-state posterior of the untested node according to equation (4).

[0115] In this embodiment, FD002 sets variables and sensor measurements for condition identification, node feature construction, and online diagnostic input; for the tasks of fc and ft fault classification and vc, vb, and vt three-state inference, 360 synthetic supervision samples with class overlap are constructed for constrained GNN training.

[0116] Synthetic samples were generated by combining six operating conditions (c1-c6) and two types of faults (fc, ft). Each "operating condition-fault" combination contained 30 samples, with 10 samples each for the early, middle, and late degradation stages. A fixed random seed (seed=42) was used, with 24 samples per group for training and 6 for validation, resulting in a total of 288 training samples and 72 validation samples. The structural constraints remained consistent with online inference: under c2, vc→vt was shielded, while the remaining 7 basic edges were enabled; under other operating conditions, all 8 basic edges were enabled, and the edge sign matrix S remained fixed.

[0117] To ensure both distinguishability between the two types of faults and to retain some state overlap, the probabilities of P30=-1, T30=+1, and T50=+1 in the fc samples increase with the degradation stage. Specifically, the target state probabilities of P30 and T30 in the three stages are 0.55, 0.75, and 0.90, respectively, and those of T50 are 0.45, 0.70, and 0.85, respectively. The ft samples primarily use the state combinations P30=0, T30=0, and T50=+1, with the probabilities of P30 and T30 being 0 being 0.80, 0.70, and 0.60, respectively, and the probabilities of T50 being +1 being 0.60, 0.80, and 0.93, respectively. A small number of adjacent states and reverse perturbations are also included.

[0118] Continuous features are generated based on node states and corresponding dead zone thresholds δs. When the state is 0, the window-averaged standardized residual is sampled within [-0.80δs, +0.80δs]; when the state is ±1, sampling is performed outside the corresponding sign direction threshold, and random perturbations are superimposed to form features such as the current standardized residual, trend, window standard deviation, and effective period ratio. For untested nodes, the probabilities of vc, vb, and vt being -1 in the fc sample are 0.82, 0.93, and 0.98 respectively with the degradation stage; the ft sample has vc=0, vb=0, vt=-1 as the main state combination, with a small amount of random perturbations.

[0119] During training, fault classification uses smoothed cross-entropy with a 0.10 label, node states use three-state cross-entropy, and node state loss is not calculated for locations lacking reliable labels. The joint loss is... L=L fault +0.65L state The training run consisted of 80 rounds. In the 80th round, during the validation phase, the fault classification accuracy was 83.33%, and the node three-state accuracy was 91.67%, indicating that the constructed features and constrained propagation model possess the corresponding classification and state inference capabilities. After training, the model parameters were fixed, and only forward inference was performed on the FD002 diagnostic window; retraining was not performed for individual samples. The network implementation parameters are shown in Table 4.

[0120] Table 4 For the diagnostic window of Unit 4 X 40:69 The aforementioned h 69 (0) , G 69 , M 69 The graph learning analysis engine, trained with input S, performs a constrained forward inference to obtain the raw fault probability: p ( fc ) = 0.24940196 ； p ( ft ) = 0.75059795 。

[0121] Simultaneously obtain vc, vb, vt The three-state posteriors of the untested nodes are shown in Table 5.

[0122] Table 5 (v) Node state uncertainty and presumed gating The collaborative controller receives the three-state posterior of the untested nodes from the graph learning analysis engine. When a unique maximum posterior state exists, it takes the maximum posterior state as the state to be predicted and its corresponding maximum posterior probability as the confidence level. At the same time, it calculates the node state uncertainty according to equation (5). U q The more concentrated the posterior, the better. U q The closer to 0, the closer the three-state posterior is. U q The closer it is to 1.

[0123] This embodiment uses a confidence threshold. θconf =0.80, uncertainty threshold θ U =0.50 and effective lifespan TTL =5. Only if the maximum a posteriori state is unique and the maximum a posteriori probability is not less than 5. θ conf and U q Not higher than θ U At that time, a valid presumed node state record is generated. For records that pass the gating, the knowledge-side evidence weight is calculated according to equation (6b): W q = w e (1-U q )。 In this embodiment, the presumed evidence benchmark weight is taken. w e =0.8. The calculation results are shown in Table 6.

[0124] Table 6 The calculation results show that, vc, vb, vt All three have a unique maximum a posteriori state, with confidence levels of 0.68517792, 0.69514048, and 0.91286445, respectively. The first two are below 0.80; corresponding to... U q The values ​​are 0.69403523, 0.66340886, and 0.31096646, respectively. Only... vt The confidence level is not less than 0.80 and U q The value is no higher than 0.50, so only vt=-1 passes the gating and forms a valid presumed node state record, while vc and vb remain in an untested state.

[0125] Within the same inference cycle, the aforementioned valid inference states are only fed back to the symbolic graph knowledge engine to participate in the knowledge-side symbolic compatibility re-inference, and are not written back to the GNN input of the current cycle. h 69 (0) It does not change what has already been generated. G 69 , M 69 And S; the subsequent knowledge-side results are only used to correct the original probability of the fault in this cycle and do not trigger the second GNN forward propagation.

[0126] (vi) Further reasoning on symbol compatibility The symbolic graph knowledge engine receives the aforementioned measured node states P30=0, T30=0, and T50=+1, as well as the valid estimated node states returned by the cooperative controller. vt =-1, and in the current valid subgraph G 69 Further reasoning on symbolic compatibility is then performed. The baseline weights for the measured node evidence are determined. w m =1, the presumed node adopts the evidence weight after uncertainty reduction. W q .

[0127] This embodiment is set N min =2、 W min =2、 N min m =2、 C min =0.80、 Q min =0.80. The symbolic graph knowledge engine sequentially performs structural reachability, sufficient evidence, and symbolic compatibility checks on candidate faults, and forms a corresponding set of candidate faults.

[0128] For candidate faults fc In the current valid subgraph G 69 The evidence set in is O c ={P30,T30,T50, vt The weights of the three measured nodes are all 1. vt The presumed evidence weight is 0.55122683, the total evidence weight is 3.55122683, the number of evidence nodes is 4, and the number of actual measured nodes is 3, which meets the condition of sufficient evidence.

[0129] according to G 69 Effective propagation paths and edge symbols in the text. fc The predicted state signs for the above nodes are P30=-1, T30=+1, and T50=+1, respectively. vt =-1. Although c2 crosses the layer edge. vc→vt Close, but fc It is still possible to follow fc→vc→vb→vt spread to vt Compared with actual evidence, T50 = +1 and vt =-1 is compatible, P30=0 and T30=0 are incompatible.

[0130] Calculated according to equations (6c), (6d), and (6e) C c=0.43681435、 E c =0.88780671、 Q c =0.38780671. Because... C c < C min and Q c < Q min Therefore fc Determined as a candidate with incompatible symbols, proceed to... F X .

[0131] For candidate faults ft The evidence set is O t ={T50, vt The T50 weight is 1. vt The weight is 0.55122683, and the total evidence weight is 1.55122683. Although T50 = +1 and vt =-1 are all with ft The predicted states are compatible, but the total weight of evidence is lower than that of the predicted states. W min =2, and the number of measured nodes is 1, which is less than N min m =2, therefore the evidence is insufficient. ft Enter F U .

[0132] This week fc, ft All passed prior screening and in G 69 The intermediate structure is reachable, ultimately yielding: F= { fc, ft} ；F P = { fc, ft} ；F R = { fc, ft} ；F U = { ft} ；F C =∅；F X = { fc}

[0133] The candidate judgment results are shown in Table 7. The symbolic graph knowledge engine returns the candidate fault set, evidence sufficiency, and symbol compatibility results to the collaborative controller.

[0134] (vii) Fault score correction, fusion uncertainty and diagnostic decision The collaborative controller receives the candidate fault set and symbol compatibility results returned by the symbolic graph knowledge engine, and combines them with the original fault probability p(F) output by the graph learning analysis engine. k First, the knowledge consistency index K is calculated according to equation (7), and then graded correction is performed according to equation (7a); subsequently, the fusion uncertainty U is calculated according to equation (7b). F , will U F The maximum correction score of the candidate fault is compared to form a fusion diagnostic decision.

[0135] This week's F C =∅, meaning there are no compatible candidates, therefore it falls into F. C The sum of the original probabilities of the faults is 0, resulting in a knowledge consistency index K=0. This embodiment sets a knowledge consistency alarm threshold θ. K =0.30, model high confidence threshold θp=0.80. The aforementioned graph learning analysis engine outputs p(fc)=0.24940196, p( ft )=0.75059795, the maximum original probability of failure is 0.75059795<θp=0.80, therefore, although K<θK, the "low knowledge consistency / high model confidence" diagnostic label will not be generated in this cycle.

[0136] In this embodiment, λ1=0.80, λ2=0.50, and λ3=0.20. Based on the candidate fault set assignment, fc ∈F X This belongs to the sign-incompatible branch, and its original fault probability is reduced by λ3 to obtain: p′( fc =0.20 × 0.24940196 = 0.04988039.

[0137] ft∈FU, belonging to the insufficient evidence branch, its corrected score remains unchanged with the original probability of the fault, and an "insufficient evidence" diagnostic label is added: p′(ft)=0.75059795.

[0138] The sum of the correction scores for the two candidates is approximately 0.80047835. According to equation (7b), the fusion uncertainty for this period is: U F =0.19952165.

[0139] Therefore, the fusion fault score for this cycle is fc =0.04988039、 ft =0.75059795, fusion uncertainty U F =0.19952165; the sum of the three is 1 within the range of numerical rounding error.

[0140] This week ft The fusion fault score of 0.75059795 is higher than that of U. F =0.19952165, therefore... ft This has been identified as the current preferred candidate fault. Because... ft ∈F U Output " ft Status: "Insufficient evidence / Pending confirmation"; fc ∈F X ,reserve" fc The "symbol incompatibility" flag constitutes the fusion diagnostic result for this cycle.

[0141] (viii) Synchronous output of digital twin The collaborative controller summarizes the diagnostic information generated by each module in this cycle, encapsulates it into a fusion diagnostic result, and sends it to the twin model through the twin state synchronization interface.

[0142] The input condition for this period is c2, corresponding to a current valid subgraph containing 7 enabling edges; the measured node states are {P30=0, T30=0, T50=+1}; after uncertainty gating, the valid inferred node state {vt=-1} is obtained, carrying the corresponding confidence level, node state uncertainty, TTL, and generation condition identifier. Knowledge-side further reasoning yields F. U ={ ft}、F C =∅、F X ={ fc}, Knowledge consistency index K=0. After hierarchical correction, the fusion fault score is { fc =0.04988039, ft =0.75059795}, fusion uncertainty U F =0.19952165. Because... ft The fusion fault score is higher than U F ,Sure ft Select the current preferred candidate fault and append " ft "Insufficient evidence / to be confirmed" and " fc The symbol is incompatible.

[0143] The above-mentioned fusion diagnostic results are synchronized to the twin model. The twin will... ftRecord the current preferred candidate fault, and also record insufficient evidence, fusion uncertainty and pending confirmation status to complete the diagnostic status update of Unit 4, cycle 69, and realize the complete closed loop from FD002 running measurement data input, structural constraint generation, graph learning reasoning, knowledge re-reasoning to fusion diagnostic output.

[0144] Example 2: This embodiment uses the pre-calibration parameters, operating condition identification model, node mapping, and symbol graph rules from Embodiment 1, without repeating the training and parameter settings. It uses the running settings variables and sensor measurement data from the 258th cycle of Unit 1 of the FD002 test set as diagnostic inputs to illustrate the symbol compatibility, insufficient evidence, and fusion diagnostic processes.

[0145] (a) Input data and multi-source data access. The multi-source data access unit reads the 258th cycle and the previous 29 cycles of FD002 Unit 1 to form X. 229:258 Causal operation window. Each cycle includes the operation settings variables u1, u2, u3, and sensor measurement x. i The knowledge-side observation nodes are P30, T30, T50, and EPR. The 258th cycle operation is set to (10.0076, 0.2501, 100.0), which is identified as operating condition c2. At the same time, the operation window, measurement validity, and node mapping results are generated.

[0146] (II) Node Status and Graph Learning Input Formation. Data preprocessing utilizes the near-health baseline and dead zone threshold related to the operating conditions fixed during the training phase. For X... 229:258 After periodic standardization and window statistics calculation, the window-averaged standardized residuals for P30, T30, and T50 are -1.62186, 2.32007, and 2.77917, respectively. Comparing these with fixed thresholds of 0.83575, 0.82749, and 1.12076, the measured node state Ψ is obtained. obs,258 ={P30=-1, T30=+1, T50=+1}; EPR is not considered reliable knowledge evidence due to its excessively small reference fluctuation scale. Simultaneously, standardized residuals, window mean and trend, runtime setting variables, measurement validity, and node type / unmeasured identifier are encoded as initial node features h. 258 (0) The output of this stage is {c2, Ψ}. obs,258 h 258 (0)}

[0147] (III) Current structural constraint formation and bilateral application. The symbolic graph knowledge engine generates M based on working condition c2. 258 G 258And S. Where vc→vt is masked, and the remaining 7 basic edges remain enabled. The symbolic graph knowledge engine limits the candidate propagation paths on the knowledge side based on G258, M258, and S and calculates the path symbols; simultaneously, the collaborative controller sends {G258, M258, S} to the graph learning analysis engine to constrain the propagation of messages on the data side. The graph learning analysis engine uses the model trained and fixed based on 360 synthetic supervised samples from Example 1, with {h 258 (0) G 258 M 258 Given S as input, a forward inference is performed, yielding p(fc) = 0.95139027 and p(ft) = 0.04860971. vc, vb, and vt are all presumed to be -1 and pass the gating, with corresponding evidence weights of 0.74615689, 0.75009289, and 0.78734436, respectively. The above valid presumed states, along with the tested states P30, T30, and T50, enter the knowledge-side re-inference.

[0148] (iv) Symbol compatibility judgment. The symbolic graph knowledge engine in G 258 Path propagation is performed on candidate fc. From fc→vc and subsequent valid paths and edge symbols, it can be seen that the predicted states of fc for P30, T30, T50, vc, vb, and vt are -1, +1, +1, -1, -1, and -1 respectively, which are all consistent with the measured states and valid inferred states. Therefore, the evidence set O of fc is... c ={P30, T30, T50, vc, vb, vt}, with 6 evidence nodes, 3 measured nodes, and a total evidence weight of 5.28359414, satisfying the sufficiency threshold; further calculation yields C c =1.00000000, E c =0.88059902、Q c =0.88059902. Therefore, fc enters the symbol-compatible candidate set F. C This indicates that the predicted propagation direction of fc under the current structural constraints is compatible with the node state evidence formed by the operating window.

[0149] (v) Judgment of Insufficient Evidence. For candidate ft, the effective path ft→vt→T50 predicts vt=-1 and T50=+1, which is consistent with the effective inference vt=-1 and the actual T50=+1. Therefore, the evidence set O is insufficient. t ={T50, vt}. Where T50 has a weight of 1, vt has a weight of 0.78734436, and the total evidence weight is 1.78734436; however, the actual number of tested nodes is 1, and the total evidence weight is lower than W. min =2, therefore, it is not enough to confirm ft as a candidate for sign compatibility, nor can it be determined as sign incompatibility. Therefore, ft is included in the insufficient evidence set F. U .

[0150] (vi) Integration of diagnostic results. The final knowledge-side output is F. C ={fc}、F U ={ft}、F X =∅, knowledge consistency index K=0.95139029. fc and ft belong to the sign compatibility and insufficient evidence branches respectively, and the correction scores remain unchanged, i.e., p′(fc)=0.95139027 and p′(ft)=0.04860971; the fusion uncertainty U in this period F =0.0. After correction, fc is still the candidate with the highest score, and fc∈F C The collaborative controller identifies fc as the fusion diagnostic fault for this cycle. The current operating condition, effective graph structure, measured node status, effective estimated status, fusion fault score, fusion uncertainty, and candidate set assignment are all synchronized to the twin model.

[0151] Embodiments of this application also propose an electronic device, including a processor and a memory; the memory is used to store a computer program executable by the processor; the processor is used to execute the computer program in the memory to implement the digital twin diagnostic method of any of the above embodiments.

[0152] Embodiments of this application also propose a computer-readable storage medium that, when an executable computer program in the storage medium is executed by a processor, enables the implementation of the digital twin diagnostic method of any of the above embodiments.

[0153] Embodiments of this application also propose a computer program product, including a computer program that, when executed by a processor, implements the digital twin diagnostic method of any of the above embodiments.

[0154] Regarding the apparatus in the above embodiments, the specific manner in which the processor performs the operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0155] Embodiments of this application also provide an electronic device 600, such as... Figure 6 As shown, the electronic device 600 includes a memory 601 and a processor 602. The memory 601 is used to store computer programs executable by the processor 602; the processor 602 is used to execute the computer programs in the memory 601 to implement the digital twin diagnostic method provided in any of the above embodiments.

[0156] The electronic device 600 also includes a communication interface 603. The processor 602, memory 601, and communication interface 603 are connected via a communication bus and communicate with each other.

[0157] Processor 602 may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of programs in the above scheme.

[0158] Communication interface 603 is used to communicate with other devices or communication networks, such as Ethernet, Radio Access Network (RAN), Wireless Local Area Networks (WLAN), etc.

[0159] The memory 601 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory may exist independently and be connected to the processor via a bus. The memory may also be integrated with the processor.

[0160] In this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The term "multiple" refers to two or more unless otherwise expressly defined.

[0161] The above description of the embodiments is intended to enable those skilled in the art to understand and apply this application. It will be apparent to those skilled in the art that various modifications can be easily made to these embodiments, and the general principles described herein can be applied to other embodiments without creative effort. Therefore, this application is not limited to the embodiments described herein, and any improvements and modifications made by those skilled in the art based on the disclosure of this application without departing from the scope and spirit of this application are within the scope of this application.

Claims

1. A digital twin diagnostic system, characterized in that, The digital twin diagnostic system includes: a multi-source data access unit, a symbolic graph knowledge engine, a graph learning and analysis engine, a mechanism data collaborative controller, and a twin model; The multi-source data access unit is used to generate an operation data sequence based on the received sensor measurement data, operation / control data, and operating condition characterization data, and send the operation data sequence to the symbolic graph knowledge engine and the graph learning analysis engine respectively. The operation data sequence includes at least: current operating condition, operation / control data, valid measured data, measured node status, and unmeasured node identifier. The symbolic graph knowledge engine is used to generate a currently valid subgraph G based on a multi-condition symbolic directed graph knowledge model, according to the current operating condition. t Dynamic edge mask matrix M t The structural constraints, consisting of the edge symbol matrix S, are sent to the mechanism data collaborative controller; the mechanism data collaborative controller is used to distribute the structural constraints to the graph learning analysis engine. The graph learning analysis engine is used to perform reasoning based on the running data sequence and the structural constraints, output the original failure probability and the three-state posterior of each untested node, and send the original failure probability and the three-state posterior of the untested node to the mechanism data collaborative controller. The mechanism data collaborative controller is used to calculate the uncertainty of node state based on the three-state posterior of untested nodes, filter out valid inferred node state records based on the node state uncertainty, unique maximum posterior probability, confidence threshold, and uncertainty threshold, and send the valid inferred node state records back to the symbolic graph knowledge engine. The symbolic graph knowledge engine is also used to determine the measured node state, the valid inferred node state record, and the current valid subgraph. The system performs evidence weighting, structural reachability judgment, and symbol compatibility reasoning to obtain a candidate fault set consisting of multiple candidate faults, and sends it to the mechanism data collaborative controller. The mechanism data collaborative controller is further configured to perform hierarchical correction on the original probability of the fault according to the affiliation of each candidate fault in the candidate fault set, obtain the candidate fault correction score of all candidate faults and calculate the fusion uncertainty, generate the fusion diagnosis result, and send the fusion diagnosis result to the twin model. The twin model is used to receive and display the fusion diagnostic results.

2. The digital twin diagnostic system according to claim 1, characterized in that, The symbolic graph knowledge engine is further configured to establish a predetermined correspondence between the state variable nodes, operation / control variable nodes, and fault event nodes in the multi-condition symbolic directed graph knowledge model and the corresponding nodes in the graph learning and analysis engine; and to establish message propagation edges in the graph learning and analysis engine based on the predetermined correspondence and the directed edge constraints in the multi-condition symbolic directed graph knowledge model, so that the causal edges in the multi-condition symbolic directed graph knowledge model correspond to the message propagation edges in the graph learning and analysis engine.

3. The digital twin diagnostic system according to claim 2, characterized in that, The reasoning based on the running data sequence and the structural constraints includes: In the message generation phase of graph learning message passing, the graph learning analysis engine uses the dynamic edge mask matrix M... t The edge symbol matrix S and graph attention weights constrain message propagation, wherein the dynamic edge mask matrix M... t The edge symbol matrix S is used to determine whether the corresponding directed edge is allowed to participate in message propagation. The edge symbol matrix S is used to determine the same-direction or opposite-direction feature transformation adopted by the directed edge that is allowed to participate in message propagation. The graph attention weight is used to adjust the propagation strength of messages that are allowed to propagate through the dynamic edge mask matrix and whose feature transformation type has been determined according to the edge symbol matrix. At the current inference moment t The l In layer message passing, by node i To the node j The message being transmitted is determined based on the following conditions: in, For the current reasoning moment No. The message passed from node i to node j in the layer. For node i at the th Layer node characteristics, The dynamic side mask matrix The middle corresponds to the current inference time. Directed edge from node i to node j matrix elements, Based on the edge symbol Choose the same or opposite feature transformation, For directed edges The corresponding edge symbol, For graph attention weights, Determined based on the following conditions: in, For the first The positive edge transformation matrix of the layer, For the first Negative edge transformation matrix of the layer.

4. The digital twin diagnostic system according to claim 1, characterized in that, The process of filtering for valid inferred node state records based on the node state uncertainty, unique maximum posterior probability, confidence threshold, and uncertainty threshold includes: The one with the highest probability among the three states of the untested node is taken as the unique maximum posterior probability. The unique maximum posterior probability is determined to be greater than or equal to the confidence threshold, and the node state uncertainty is less than or equal to the uncertainty threshold. Generate valid presumed node state records.

5. The digital twin diagnostic system according to claim 1, characterized in that, The execution of evidence weighting, structural reachability judgment, and symbolic compatibility reasoning yields a candidate fault set consisting of multiple candidate faults, including: By performing evidence weighting, structural reachability judgment, and symbolic compatibility reasoning, we obtain evidence sufficiency, symbolic compatibility results, and candidate fault scores. Based on the evidence sufficiency, symbolic compatibility results, and candidate fault scores, we obtain a candidate fault set consisting of multiple candidate faults. Candidate Fault The sign compatibility result is calculated according to the following formula: in, The actual state symbol of the observation node q. Candidate fault f k Predicted state symbol for the node, symbol compatibility results C k It reflects the degree of consistency between the predicted state symbol and the actual state symbol. Candidate Fault Sufficientness of evidence and candidate fault score Calculate according to the following formulas respectively: in, Candidate fault The corresponding set of observation nodes For the set of observation nodes Number of nodes Let q be the evidence weight. The evidence benchmark weights for the measured nodes.

6. The digital twin diagnostic system according to claim 1, characterized in that, The candidate fault set includes a priori candidate set, a structurally reachable candidate set, a candidate set with insufficient evidence, and a candidate set with incompatible symbols. The original fault probabilities are subjected to hierarchical correction based on the following conditions: in, , For the prior candidate set, For structurally reachable candidate sets, This is a candidate set with insufficient evidence. For the candidate set of symbol incompatibility, The candidate fault correction score is the value for the k-th candidate fault. Let k be the original probability of the k-th candidate fault, where k is a positive integer.

7. The digital twin diagnostic system according to claim 6, characterized in that, The calculation of fusion uncertainty includes: The sum of the candidate fault correction scores for all candidate faults is calculated. The difference between 1 and the sum is taken as the fusion uncertainty; The maximum candidate fault correction score among all candidate fault correction scores is compared with the fusion uncertainty. If the fusion uncertainty is less than the maximum candidate fault correction score, the candidate fault corresponding to the maximum candidate fault correction score shall be selected as the preferred candidate fault.

8. The digital twin diagnostic system according to claim 7, characterized in that, The candidate fault set also includes a symbol-compatible candidate set, and the calculation of the fusion uncertainty further includes: If the preferred candidate fault belongs to the symbol compatible candidate set, the preferred candidate fault is determined as the current fusion diagnosis fault, and the diagnosis status is marked as compatible. If the preferred candidate fault belongs to the insufficient evidence candidate set, the diagnostic status is marked as insufficient evidence, pending confirmation. If the preferred candidate fault belongs to the symbol incompatible candidate set, or does not belong to the prior candidate set, or belongs to the prior candidate set but does not belong to the structure reachable candidate set, the preferred candidate fault is taken as the result to be reviewed. The proportion of the probability that falls into the symbol-compatible candidate set among all original fault probabilities is defined as the knowledge consistency index K. If the knowledge consistency index K is less than the knowledge consistency threshold θK, and the maximum original fault probability among all original fault probabilities is higher than the model high confidence threshold θp, a low knowledge consistency / model high confidence label is added, and the relevant node and directed edge information is recorded for subsequent verification.

9. A digital twin diagnostic method, characterized in that, The digital twin diagnostic method includes: Based on the received sensor measurement data, operation / control data, and operating condition characterization data, an operation data sequence is generated. The operation data sequence includes at least: current operating condition, operation / control data, valid measured data, measured node status, and unmeasured node identifier. Based on a multi-condition symbolic directed graph knowledge model, a subgraph is generated according to the current operating condition, consisting of the currently valid subgraph. Dynamic edge mask matrix The structural constraints consist of the edge symbol matrix S; Based on the operational data sequence and the structural constraints, reasoning is performed to output the original probability of the fault and the three-state posterior of each untested node; The uncertainty of node state is calculated based on the three-state posterior of untested node. Valid inferred node state records are obtained by filtering based on the node state uncertainty, unique maximum posterior probability, confidence threshold, and uncertainty threshold. Based on the measured node status, the valid estimated node status record, and the current valid subgraph By performing evidence weighting, structural reachability judgment, and symbol compatibility reasoning, a candidate fault set consisting of multiple candidate faults is obtained; Based on the affiliation of each candidate fault in the candidate fault set, a hierarchical correction is performed on the original probability of the fault to obtain the candidate fault correction score of all candidate faults and calculate the fusion uncertainty to generate the fusion diagnosis result. Receive and display the fusion diagnostic results.

10. The digital twin diagnostic method according to claim 9, characterized in that, The dynamic side mask matrix M t It is the current inference moment. Used to label each directed edge A binary matrix indicating whether participation in message propagation / path filtering is allowed; the effective subgraph G t In the current inference moment Based on the current operating conditions The dynamic edge mask matrix M is used to define the multi-condition symbolic directed graph knowledge model. t The remaining subgraph after removing the masked directed edges; the edge symbol matrix. It stores each directed edge A matrix of qualitative causal influence symbols; and, =[ ,in, For dynamic edge mask matrix, To correspond to the current inference moment Directed edge from node i to node j matrix elements, Determined based on the following conditions: in, For the current reasoning moment The current operating conditions Current operating conditions Enable condition; The dynamic edge mask matrix By applying the multi-condition symbolic directed graph knowledge model, the current valid subgraph is obtained. .

Images