Diagnostic protocol generation method and system based on multi-scale hypergraphs and constrained optimization

Abstract

A diagnosis scheme generation method and system based on multi-scale hypergraph and constraint optimization belong to the technical field of industrial equipment health management. The method comprises the following steps: receiving fault diagnosis query information expressed in natural language, performing semantic analysis on the fault diagnosis query information to form semantic guidance information; extracting task nodes from a technical information set and constructing multi-scale nested subgraphs according to the relationships between the task nodes; based on the multi-scale nested subgraphs, dynamically reconstructing a task path, representing the task nodes as hypergraph nodes, and constructing a multi-scale hypergraph; performing structure perception regularization processing on the multi-scale hypergraph to obtain a multi-scale hypergraph; performing cross-scale attention processing on the multi-scale hypergraph to obtain a multi-scale fusion hypergraph; generating a fault diagnosis path based on the multi-scale fusion hypergraph, sorting the fault diagnosis path, and outputting a diagnosis scheme. The present application can significantly improve the intelligent level and decision efficiency of complex system fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment health management, and in particular to a diagnosis scheme generation method and system based on multi-scale hypergraph and constraint optimization. BACKGROUND

[0002] In the field of industrial equipment health management, the generation of fault diagnosis schemes mainly relies on the comprehensive use of equipment knowledge, historical data and expert experience. At present, a relatively basic and linear processing method is mainly adopted. Generally, information from different sources, such as equipment manuals, maintenance records, expert rules, etc., is established into independent, isolated knowledge bases or databases. When facing a new fault diagnosis task, keyword matching or rule-based query is mainly used to search for historical cases or preset fault handling processes directly related to the current fault symptoms in the independent knowledge base, and finally a diagnosis scheme is formed. Effective integration of fault diagnosis knowledge is the core challenge to improve the accuracy and efficiency of diagnosis. SUMMARY

[0003] To overcome the shortcomings of the prior art, the purpose of the present application is to provide a diagnosis scheme generation method and system based on multi-scale hypergraph and constraint optimization, which can significantly improve the intelligent level and decision efficiency of complex system fault diagnosis.

[0004] To achieve the purpose of the application, in a first aspect, the embodiments of the present application provide a diagnosis scheme generation method based on multi-scale hypergraph and constraint optimization, which comprises:

[0005] Receiving fault diagnosis query information expressed in natural language, performing semantic analysis on the fault diagnosis query information to form semantic guidance information;

[0006] Based on the semantic guidance information, extracting task nodes from a technical information set, and constructing multi-scale nested subgraphs according to the relationships between the task nodes;

[0007] Based on the multi-scale nested subgraphs, dynamically reconstructing the task path, representing the task nodes as hypergraph nodes, and constructing a multi-scale hypergraph;

[0008] Performing structure-aware regularization processing on the multi-scale hypergraph to obtain a regularized multi-scale hypergraph;

[0009] Performing cross-scale attention processing on the regularized multi-scale hypergraph to obtain a multi-scale fusion hypergraph;

[0010] Generating a fault diagnosis path based on the multi-scale fusion hypergraph, and sorting the fault diagnosis path, and outputting the fault diagnosis path with high ranking as a diagnosis scheme.

[0011] In a possible implementation, the semantic analysis on the fault diagnosis query information forms semantic guidance information, including:

[0012] The intent recognition and semantic decomposition are performed on the fault diagnosis query information to obtain semantic representation information;

[0013] The context semantic modeling and entity extraction processing are performed on the semantic representation information to obtain structured information;

[0014] The inference is performed on the structured information to obtain potential constraint information, and the structured information and the potential constraint information form the semantic guidance information.

[0015] In a possible implementation, the method further includes: extracting task nodes from a technical information set based on the semantic guidance information, and constructing multi-scale nested subgraphs according to relationships among the task nodes, including:

[0016] Based on the semantic guidance information and the technical information set, structured matching processing is performed to extract task nodes, and the task nodes include core diagnostic targets, related components and key working condition features;

[0017] Based on parent-child relationships and dependency relationships among the task nodes, multi-scale nested subgraphs are constructed, wherein each multi-scale nested subgraph corresponds to a task structure at a different abstraction level.

[0018] In a possible implementation, based on the multi-scale nested subgraphs, a task path is dynamically reconstructed, the task nodes are represented as hypergraph nodes, and a multi-scale hypergraph is constructed, including:

[0019] According to relationships among the task nodes in the semantic guidance information and the multi-scale nested subgraphs, a node sequence of the task path is dynamically adjusted to obtain a reconstructed task path;

[0020] The task nodes in the multi-scale nested subgraphs are mapped to hypergraph nodes, and an attribute vector of the hypergraph node includes semantic features, structural features and constraint features;

[0021] Based on the reconstructed task path and the hypergraph nodes, a multi-scale hypergraph is constructed.

[0022] In a possible implementation, the multi-scale hypergraph is subjected to structure-aware regularization processing to obtain a regularized multi-scale hypergraph, including:

[0023] The semantic similarity scores of nodes and hyperedges of the multi-scale hypergraph are calculated, for each node, a hyperedge with a high semantic similarity score is selected, and a sparse connection matrix of the multi-scale hypergraph is constructed;

[0024] Based on the sparse connection matrix, a hyperedge with a connection strength greater than or equal to a preset threshold is retained to form a pruned sparse connection matrix;

[0025] Logical verification and conflict reconciliation are performed on the pruned sparse connection matrix by constructing constraint conditions to obtain the regularized multi-scale hypergraph.

[0026] In a possible implementation, the logical verification and conflict reconciliation performed on the pruned sparse connection matrix by constructing constraint conditions to obtain the regularized multi-scale hypergraph include:

[0027] A structural constraint hypergraph is constructed, and the structural constraint hypergraph includes constraint conditions for nodes and hyperedges of the multi-scale hypergraph.

[0028] A hypergraph neural network is used to model the structural constraint hypergraph to learn node selection weights and hyperedge constraint tension, and a structural consistency optimization model is obtained.

[0029] Logical verification and conflict reconciliation are performed on the pruned sparse connection matrix based on the structural consistency optimization model to obtain the regularized multi-scale hypergraph.

[0030] In a possible implementation, the cross-scale attention processing performed on the regularized multi-scale hypergraph to obtain a multi-scale fusion hypergraph includes:

[0031] For each structural scale, semantic attention weights between each node of the multi-scale hypergraph and hyperedges connected to the node are calculated, node information is aggregated and node representations are updated based on the semantic attention weights, and a fused hypergraph node is obtained.

[0032] The hyperedges of the multi-scale hypergraph at each structural scale are linearly mapped, and the hyperedges of different structural scales are fused to obtain a fused hypergraph hyperedge.

[0033] The fused hypergraph node and the fused hypergraph hyperedge are spliced to obtain the multi-scale fusion hypergraph.

[0034] In a possible implementation, the fault diagnosis path is generated based on the multi-scale fusion hypergraph, and the fault diagnosis path is sorted, and a top-ranked fault diagnosis path is output as a diagnosis scheme, which includes:

[0035] Based on the multi-scale fusion hypergraph, the fault diagnosis query information is matched with the hypergraph nodes of the multi-scale fusion hypergraph to generate candidate diagnosis paths.

[0036] The structural consistency and semantic rationality of each candidate diagnosis path are evaluated, and the candidate diagnosis paths are sorted according to the evaluation results.

[0037] output the candidate diagnosis path with the highest ranking as a diagnosis scheme.

[0038] To achieve the object of the present application, the second aspect, the embodiments of the present application provide a system, comprising a storage medium and one or more processors; the storage medium stores a computer program, and the computer program is called by the one or more processors to implement any one of the above-mentioned diagnosis scheme generation methods based on multi-scale hypergraph and constraint optimization.

[0039] The technical scheme provided by the embodiments of the present application at least brings the following beneficial effects: The present application aims at the core problem of low accuracy and efficiency of diagnosis scheme generation caused by the isolation of knowledge bases and the dependence on keyword matching in the prior art, and realizes a technical breakthrough through multi-scale hypergraph modeling and constraint consistency optimization. The technical effects include: through semantic analysis and cross-scale attention mechanism, heterogeneous knowledge is deeply fused, breaking the "knowledge island" and realizing the qualitative change from linear retrieval to complex network reasoning; dynamic reconstruction and optimization of the diagnosis path ensure the logical rigor and global optimality of the scheme generation. The scheme significantly improves the accuracy and efficiency of fault diagnosis, can quickly generate an optimal diagnosis path that is practical and feasible, effectively reduces the operation and maintenance cost, and greatly enhances the intelligent level of industrial equipment health management. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the following drawings.

[0041] Figure 1 is a schematic flow chart of a diagnosis scheme generation method based on multi-scale hypergraph and constraint optimization provided by the embodiments of the present application. DETAILED DESCRIPTION

[0042] In order to make the object, technical scheme and advantages of the present application more clear, the following will further describe the embodiments of the present application in detail with reference to the drawings.

[0043] The embodiments of the present application provide a diagnosis scheme generation method based on multi-scale hypergraph and constraint optimization, which is described with reference to Figure 1 The method comprises the following steps:

[0044] Step one: receiving fault diagnosis query information expressed by natural language, performing semantic analysis on the fault diagnosis query information to form semantic guidance information.

[0045] The user can express the fault diagnosis query information of the device through natural language, such as describing motor bearing vibration data, environmental characteristics under operating conditions, and descriptions and detection requirements for faults.

[0046] Optionally, the semantic analysis of the fault diagnosis query information forms semantic guidance information, including the following steps:

[0047] The intent recognition and semantic decomposition of the fault diagnosis query information obtain semantic representation information; the context semantic modeling and entity extraction processing of the semantic representation information obtain structured information; the inference of the structured information obtains potential constraint information, and the structured information and the potential constraint information form the semantic guidance information.

[0048] Among them, for the fault diagnosis query information of the user, a pre-trained language model can be used for intent recognition and semantic decomposition of the input. Through context semantic modeling and entity extraction, the core diagnosis target, related components and key working condition characteristics are identified. Further, based on the extraction result, potential constraint information such as function sequence, state trigger and execution dependency is inferred, and it is constructed as a structured demand tensor and a potential constraint information set to form semantic representation information, providing semantic guidance for subsequent graph structure modeling.

[0049] This step can accurately convert fuzzy and unstructured natural language queries into machine-understandable and inferable structured semantic guidance information. Through intent recognition and semantic decomposition, the core diagnosis target, related components and key working conditions of the user are accurately captured to ensure the accuracy of the diagnosis direction. Through context reasoning, the logical relationships and potential constraints implied in the description are excavated and made explicit, and structured information rich in logic is constructed. This not only provides a high-quality semantic basis for the construction of subsequent graph networks, but also enables the entire diagnosis system to be intelligently guided based on a deep understanding of user intent, thereby ensuring the accuracy and logical rigor of the generated diagnosis scheme from the source.

[0050] Step 2: Based on the semantic guidance information, task nodes are extracted from the technical information set, and multi-scale nested subgraphs are constructed according to the relationships between the task nodes.

[0051] Optionally, this step can include: based on the semantic guidance information and the structured matching processing of the technical information set, extracting task nodes, the task nodes including core diagnosis targets, related components and key working condition characteristics; based on the parent-child relationship and dependency relationship between the task nodes, multi-scale nested subgraphs are constructed, wherein each multi-scale nested subgraph corresponds to a task structure at different abstraction levels.

[0052] The technical information set can include multi-source heterogeneous data, including but not limited to: a technology roadmap describing the hierarchical structure and dependency relationship of task execution; a fault tree analysis diagram representing the failure mode and propagation path; a diagnostic manual and document containing detailed descriptions and constraint rules of task nodes; sensor data and topology graph providing device operating status and connection relationship between functional modules.

[0053] The matching of the semantic guidance information and the technical information set can be achieved by the following steps:

[0054] The core diagnostic target, related components and key working condition features in the semantic guidance information are encoded into vector representations using a pre-trained language model, and similarity calculation is performed with the task node vectors in the technical information set to obtain a similarity score.

[0055] According to the similarity score, the task nodes related to the semantic guidance information are screened out. For example, if the query contains "abnormal vibration of motor bearing", the "bearing fault diagnosis" node in the technical information set is matched.

[0056] The latent constraint information is aligned with the constraint rules in the technical information set to ensure that the extracted task nodes meet the constraint rules.

[0057] After matching, the task nodes are extracted, which can include: core diagnostic target, such as "bearing fault diagnosis" and "motor fault positioning"; related components, such as "bearing", "motor", and "sensor"; key working condition features, such as "high load", "impact", and "sand dust". These nodes serve as the basis for subsequent multi-scale nested subgraph construction.

[0058]

[0059] For the parent-child relationship and dependency relationship between the task nodes, the step of constructing a multi-scale nested subgraph is performed. The core goal of this step is to organize the task nodes extracted in the previous step into multiple nested subgraph structures according to their inherent parent-child relationship and dependency relationship, with each subgraph corresponding to a different abstract level of task structure. This process provides a structured foundation for subsequent dynamic reconstruction of task paths and construction of multi-scale hypergraphs.

[0060] This step can include:

[0061] Based on the parent-child relationship between the task nodes, the task nodes are divided into several abstract levels, with the top-level nodes representing the overall diagnostic target, the middle-level nodes representing sub-tasks or modules, and the bottom-level nodes representing specific operations or detection units, thereby forming a hierarchical task structure.

[0062] ​Based on the dependency relationship between task nodes, the task nodes with logical association or constraint conditions are connected by constructing hyper-edges to form a dependency network, wherein each hyper-edge represents the collaborative execution or constraint relationship between multiple task nodes, thereby capturing the complex interaction between tasks.

[0063] The hierarchical task structure is combined with the dependency network to construct a multi-scale nested subgraph, wherein each nested subgraph corresponds to a task structure at an abstract level, the nested subgraphs at different levels are connected by parent-child relationship, and the logical transmission is performed across levels by the hyper-edges of the dependency relationship, thereby forming a multi-scale, nested, high-order associated task graph structure.

[0064] The parent-child relationship describes the hierarchical nested structure between task nodes and is the basis for constructing a multi-scale nested subgraph. The hierarchy can include a top layer, a middle layer, and a bottom layer. The top layer represents system-level or function-level tasks, such as "system fault diagnosis" and "device health management". These nodes are usually the entry of the entire diagnostic task. The middle layer represents module-level or subsystem-level tasks, such as "motor fault positioning" and "sensor fault diagnosis". These nodes are subtasks of the top-level tasks. The bottom layer represents operation-level or atomic-level tasks, such as "bearing fault diagnosis" and "motor control unit fault diagnosis". These nodes are further decompositions of the middle-level tasks.

[0065] The parent-child relationship is mapped. The parent-child relationship is represented by a cross-layer connection edge set. For example, "bearing fault diagnosis" is a child node of "motor fault positioning". "Motor fault positioning" is a child node of "system fault diagnosis".

[0066] The hierarchy is aggregated, and the fine-grained task nodes are aggregated into coarse-grained semantic units using an aggregation function. For example, "bearing fault diagnosis" and "motor control unit fault diagnosis" at the bottom layer are aggregated into "motor fault positioning" at the middle layer. "Motor fault positioning" and "sensor fault diagnosis" at the middle layer are aggregated into "system fault diagnosis" at the top layer.

[0067] Based on the dependency relationship between task nodes, a super edge is constructed to connect task nodes with logical association or constraint conditions, and then a dependency network is formed. In this process, the "dependency relationship" refers to the logical connection existing in the execution process of the task nodes, such as the order, conditional constraints or data transmission, which determines the coordination mode and execution order between the task nodes. The "super edge" is a special data structure that can connect multiple task nodes and represent the collaborative execution or constraint relationship between these nodes. Each super edge not only records the connected task nodes, but also contains the specific dependency type and constraint conditions between these nodes. In the construction of the dependency network, the dependency relationship between the task nodes is first analyzed and extracted, and then according to the extraction results, the task nodes with logical association or constraint conditions are connected through the super edge. In this way, the originally dispersed task nodes are organized into an ordered and constrained dependency network, so that the complex interaction between tasks can be more accurately captured.

[0068] Step three: based on the multi-scale nested subgraph, dynamically reconstructing the task path, representing the task nodes as supergraph nodes, and constructing a multi-scale supergraph.

[0069] Optionally, this step includes:

[0070] According to the relationship between the semantic guidance information and the task nodes in the multi-scale nested subgraph, the node sequence of the task path is dynamically adjusted to obtain a reconstructed task path; the task nodes in the multi-scale nested subgraph are mapped to supergraph nodes, and the attribute vector of the supergraph node includes semantic features, structural features and constraint features; and a multi-scale supergraph is constructed based on the reconstructed task path and the supergraph nodes.

[0071] In an optional embodiment, according to the relationship between the semantic guidance information and the task nodes in the multi-scale nested subgraph, the node sequence of the task path is dynamically adjusted to obtain a reconstructed task path, which can include:

[0072] Based on the semantic guidance information, the task nodes in the multi-scale nested subgraph are semantically matched and weighted, and the relevance of each node to the user query is determined. Preferably, the semantic guidance information is matched with the task nodes in the multi-scale nested subgraph, and the relevance score of each node to the query is calculated. According to the score, each node is assigned a weight. This process not only improves the accuracy of task node screening, but also provides a quantitative basis for subsequent path dynamic adjustment.

[0073] The parent-child relationship and dependency relationship between the task nodes are combined to prioritize the nodes and dynamically adjust the node execution order. Preferably, based on the node weight distribution, the system further combines the parent-child relationship and dependency relationship between the task nodes, and can perform the prioritization by constructing a task priority ordering model. The model uses a topological sorting algorithm to ensure that the parent node is executed before the child node, while considering the influence of the dependency relationship on the node order. For example, if a node depends on the output result of another node, the dependent node must be placed in the pre-order position. By dynamically adjusting the node execution order, the system can generate a preliminary task path that is logically reasonable and efficient. This step combines the static node relationship with dynamic priority adjustment, significantly improving the flexibility and adaptability of the task path.

[0074] The task path is reconstructed and optimized through a dynamic programming and feedback mechanism to generate an optimal execution sequence. Preferably, in order to further improve the execution efficiency and accuracy of the task path, a dynamic programming algorithm is introduced in this step to find the optimal execution sequence and minimize the execution time or resource consumption. At the same time, through the feedback mechanism, the path execution effect is monitored in real time, and if it is found that a node is executing abnormally or inefficiently, the subsequent node order or the node is automatically adjusted. This dynamic reconstruction mechanism not only improves the robustness of the path, but also adapts to changes in task requirements in different scenarios, reflecting the adaptive ability of the system in complex industrial environments.

[0075] Preferably, the task nodes in the multi-scale nested subgraph are mapped to hypergraph nodes, and the attribute vector of the hypergraph node includes semantic features, structural features, and constraint features, which can include:

[0076] The task nodes in the multi-scale nested subgraph are mapped to hypergraph nodes, and the original structure and semantic information are preserved. For example, the parent node still maintains the control relationship over the child node in the hypergraph, and the dependency relationship is connected through the hyperedge. This mapping process not only simplifies the expression of complex subgraphs, but also lays a foundation for efficient computation and reasoning of hypergraphs in the future, embodying the innovation of multi-scale structure transformation to hypergraph.

[0077] An attribute vector is constructed for each hypergraph node, including semantic features, structural features, and constraint features, to fully represent the node characteristics.

[0078] The attribute vector expression of the supergraph node is optimized through feature fusion and normalization processing. In this step, feature fusion and normalization technology are used to optimize the vector. Feature fusion effectively integrates semantic, structural and constraint features through weighted averaging or neural network model to form a comprehensive feature vector. Normalization processing can eliminate the dimensional difference between different features, improve the consistency and comparability of the vector. The optimized attribute vector not only enhances the discrimination of the node in the supergraph, but also improves the learning efficiency and reasoning accuracy of the subsequent graph neural network, reflecting the important role of feature engineering in supergraph construction.

[0079] Preferably, based on the reconstructed task path and the supergraph node, a multi-scale supergraph is constructed, which can include:

[0080] According to the dependency relationship between the reconstructed task path and the supergraph node, a superedge is constructed to connect nodes with cooperative or constraint relationship. The superedge not only connects two nodes, but also connects multiple nodes at the same time, forming a high-order dependency relationship.

[0081] The supergraph nodes are divided into multiple scales, and the supergraph nodes and superedges are optimized through a graph neural network to construct a multi-scale supergraph. After the superedge is constructed, the system performs multi-scale hierarchical division on the supergraph to form a task structure at different abstraction levels. The nodes at different levels can be identified and optimized by a graph neural network to form a nested supergraph structure. The graph neural network learns the features of nodes and superedges on the supergraph through a message passing mechanism, and optimizes the structure and parameters of the supergraph.

[0082] Step four: performing structure-aware regularization processing on the multi-scale supergraph to obtain a regularized multi-scale supergraph.

[0083] Preferably, this step can include:

[0084] Calculate the semantic similarity score of the nodes and superedges of the multi-scale supergraph, select the superedges with the highest semantic similarity score for each node, and construct a sparse connection matrix of the multi-scale supergraph; based on the sparse connection matrix, retain the superedges with connection strength greater than or equal to a preset threshold to form a pruned sparse connection matrix; perform logical verification and conflict reconciliation on the pruned sparse connection matrix by constructing constraint conditions to obtain the regularized multi-scale supergraph.

[0085] Optionally, calculating the semantic similarity score of the nodes and superedges of the multi-scale supergraph, selecting the superedges with the highest semantic similarity score for each node, and constructing a sparse connection matrix of the multi-scale supergraph can include:

[0086] For each node, a semantic similarity score between the node and all hyper-edges is calculated based on the semantic embedding vector of the node and the semantic embedding vectors of all hyper-edges; for each node, the semantic similarity scores are ranked from high to low, and a number of hyper-edges with high scores are selected as candidate connection hyper-edges of the node; based on the correspondence between each node and its candidate connection hyper-edges, a bipartite graph structure of nodes to hyper-edges is constructed, and an initial sparse connection matrix is generated.

[0087] Optionally, based on the sparse connection matrix, hyper-edges with connection strength greater than or equal to a preset threshold are retained to form a pruned sparse connection matrix, which can include:

[0088] Each hyper-edge in the sparse connection matrix is traversed, and the connection strength of the hyper-edge is calculated, which is the weighted sum of the semantic similarity scores between the nodes and the hyper-edges; the connection strength of each hyper-edge is compared with a preset threshold, if the connection strength is greater than or equal to the threshold, the hyper-edge is retained, otherwise it is removed; the sparse connection matrix is updated according to the retained hyper-edges to form a pruned sparse connection matrix.

[0089] Optionally, the pruned sparse connection matrix is logically verified and conflict reconciled by constructing constraint conditions to obtain the regularized multi-scale hypergraph, which can include:

[0090] A structure constraint hypergraph is constructed, which includes constraint conditions for nodes and hyper-edges of the multi-scale hypergraph; a hypergraph neural network is used to model the structure constraint hypergraph to learn node selection weights and hyper-edge constraint tension, and a structure consistency optimization model is obtained; the pruned sparse connection matrix is logically verified and conflict reconciled based on the structure consistency optimization model to obtain a regularized multi-scale hypergraph.

[0091] The structure constraint hypergraph includes constraint conditions for nodes and hyper-edges of the multi-scale hypergraph, which can include:

[0092] Based on domain knowledge or historical diagnosis data, logical dependency relationships between nodes, semantic coverage ranges of hyper-edges, and topological constraint rules between nodes and hyper-edges are defined;

[0093] The logical dependency relationships, semantic coverage ranges, and topological constraint rules are mapped to constraint edges in the hypergraph to form an initial structure constraint hypergraph;

[0094] The initial structure constraint hypergraph is subjected to redundant constraint elimination and priority sorting to obtain a simplified structure constraint hypergraph.

[0095] By introducing domain knowledge and historical diagnosis data, the simplified structure constraint hypergraph constructed by the scheme provides high signal-to-noise ratio input for subsequent modeling, ensuring that the system has stronger reasoning ability and robustness in complex diagnosis scenarios.

[0096] The modeling of the structure constraint hypergraph by the hypergraph neural network, learning of node selection weights and hyperedge constraint tension, and obtaining of the structure consistency optimization model can include:

[0097] The structure constraint hypergraph is input into the hypergraph neural network, and the interaction relationship between nodes and hyperedges is encoded through a message passing mechanism to generate joint embedding representation of nodes and hyperedges.

[0098] Based on the joint embedding representation, the selection weight of each node and the constraint tension of each hyperedge are calculated, wherein the node selection weight reflects its importance in the diagnosis path, and the hyperedge constraint tension reflects its constraint strength on node connection.

[0099] The loss function is used to jointly optimize the node selection weight and the hyperedge constraint tension to obtain the structure consistency optimization model for guiding subsequent logical verification and conflict resolution.

[0100] The structure consistency optimization model obtained by the scheme not only improves the modeling accuracy, but also provides a reliable mathematical basis for subsequent logical verification and conflict resolution, significantly enhancing the generalization ability and stability of the model.

[0101] The logical verification and conflict resolution of the pruned sparse connection matrix based on the structure consistency optimization model to obtain the regularized multi-scale hypergraph can include:

[0102] The structure consistency optimization model is used to perform logical verification on the pruned sparse connection matrix to identify node-to-hyperedge connection pairs that violate the constraint condition.

[0103] The conflict resolution of the identified conflict connection pairs includes reallocating node selection weights, adjusting hyperedge constraint tension, or removing redundant connections to eliminate logical conflicts.

[0104] The sparse connection matrix is updated according to the resolution result, and the regularized multi-scale hypergraph is output.

[0105] This solution, based on a structural consistency optimization model, performs logical verification and conflict reconciliation on the pruned sparse connection matrix, efficiently identifying and handling node-to-hyperedge connection pairs that violate constraints. By reallocating node weights, adjusting hyperedge constraint tension, or eliminating redundant connections, the system effectively eliminates logical conflicts, ensuring the structural consistency and semantic rationality of the multi-scale hypergraph. The final output regularized multi-scale hypergraph not only improves the accuracy and reliability of inference paths but also provides high-quality structured support for intelligent decision-making in complex scenarios, possessing significant engineering application value.

[0106] In a preferred embodiment, during the construction of the structural constraint hypergraph, this application utilizes a structure-aware constraint regularization mechanism to enhance the model's ability to cluster similar semantic structures and to distinguish structural conflicts in paths. This mechanism first defines the logical dependencies between nodes, the semantic coverage of hyperedges, and the topological constraint rules between nodes and hyperedges based on domain knowledge or historical diagnostic data, mapping these to constraint edges in the hypergraph to form an initial structural constraint hypergraph. Subsequently, the initial structural constraint hypergraph is simplified through redundant constraint removal and priority ranking mechanisms.

[0107] When constructing a structurally constrained hypergraph, node consistency constraints are used to ensure semantic consistency between nodes and hyperedges. The node consistency constraint is defined as follows:

[0108] ,

[0109] in, This represents node consistency constraints. Let represent the embedding representation of the i-th task node in layer s. This represents the embedding representation of the j-th superedge. Indicates the relationship with the hyperedge The set of all connected nodes This represents the total number of superedges in the s-th layer. This represents semantic distance. This constraint encourages task nodes connected by the same hyperedge to have similar semantic representations, preventing semantically heterogeneous nodes from being incorrectly aggregated.

[0110] In logic verification and conflict reconciliation, hyperedge heterogeneity constraints are used to enhance the similarity and differences between hyperedges. The hyperedge heterogeneity constraint is defined as follows:

[0111] ,

[0112] in, For hyperedge heterogeneity constraints, This represents the total number of hyperedges at the s-th scale. Used to represent the relationship between hyperedge i and hyperedge j; is the Euclidean distance of hyperedge i and hyperedge j at the s-th scale; is a truncated linear function to limit the upper bound of distance difference, is a threshold value.

[0113] In the structure self-updating regularization term construction stage, a feedback adjustment mechanism is introduced to dynamically adjust the connection weights of nodes and hyperedges to adapt to the changes in data distribution. The specific steps include: first, based on the historical connection state and the current input data, the connection stability score of nodes and hyperedges is calculated; second, using the connection stability score, a structure self-updating regularization term is constructed to constrain the change amplitude of the connection weights of nodes and hyperedges; finally, the structure self-updating regularization term is combined with the node constraint term and the hyperedge constraint term to form a structure consistency optimization model to prevent frequent oscillation of the graph structure.

[0114] Step five: performing cross-scale attention processing on the regularized multi-scale hypergraph to obtain a multi-scale fusion hypergraph.

[0115] Optionally, this step can include:

[0116] For each structural scale, the semantic attention weight between each node of the multi-scale hypergraph and the hyperedge connected to the node is calculated, the node information is aggregated and the node representation is updated based on the semantic attention weight, and a fused hypergraph node is obtained;

[0117] The hyperedges of the multi-scale hypergraph at each structural scale are linearly mapped, and the hyperedges of different structural scales are fused to obtain a fused hypergraph hyperedge.

[0118] The fused hypergraph node and the fused hypergraph hyperedge are spliced to obtain the multi-scale fusion hypergraph.

[0119] This scheme realizes the deep fusion of nodes and hyperedges in the multi-scale hypergraph through cross-scale attention processing, and the specific technical effects can include: through the attention mechanism, the nodes can adaptively capture the semantic association between the nodes and the hyperedges, enhancing the modeling ability of complex structural relationships; through linear mapping and weighted fusion, the hyperedge information of different structural scales is effectively integrated, eliminating the semantic deviation between scales and improving the generalization ability of the model; through the splicing and integration of nodes and hyperedges, a multi-scale fusion hypergraph with complete structure and consistent semantics is formed, providing high-quality structured input for subsequent tasks.

[0120] For each structural scale, the semantic attention weight between each node of the multi-scale hypergraph and the hyperedge connected to the node is calculated, the node information is aggregated and the node representation is updated based on the semantic attention weight, and a fused hypergraph node is obtained, which can include:

[0121] For each structural scale, initialize the attention weight calculation parameters between nodes and hyper-edges.

[0122] Based on the node embedding vector and the hyper-edge feature vector, the semantic correlation score between each node and its connected hyper-edge is calculated using dot product attention or multi-head attention mechanism, and the semantic attention weight is obtained through normalization processing.

[0123] Using the calculated semantic attention weight, the hyper-edge information connected to the node is weighted and aggregated, the aggregated information is spliced or weighted with the original node representation, and the node representation is updated through a nonlinear activation function to obtain the fused hypergraph node.

[0124] Among them, the linear mapping of the hyper-edge of the multi-scale hypergraph under each structural scale is carried out, and the hyper-edges of different structural scales are fused to obtain the fused hypergraph hyper-edge, which can include:

[0125] For the hyper-edge of each structural scale, a linear mapping model is constructed to map the high-dimensional hyper-edge feature to a unified low-dimensional semantic space.

[0126] The features of the hyper-edges of different structural scales are aligned, and the aligned hyper-edge features are fused to generate hyper-edge representations with cross-scale consistency using weighted averaging or max pooling methods.

[0127] The fused hyper-edge representation is subjected to nonlinear transformation and normalization processing, and the fused hypergraph hyper-edge is output.

[0128] Among them, the fused hypergraph node and the fused hyper-edge hyper-edge are spliced to obtain a multi-scale fusion hypergraph, which can include:

[0129] The dimensions of the fused hypergraph node and hyper-edge representation are aligned, the aligned node representation and hyper-edge representation are spliced by dimension to form a unified node-hyper-edge joint representation matrix, the structural association information between nodes and hyper-edges is retained, and the spliced joint representation matrix is post-processed. The post-processing can be normalization, sparsification or dimensionality reduction operation to generate a multi-scale fusion hypergraph.

[0130] Step six: generate a fault diagnosis path based on the multi-scale fusion hypergraph, and sort the fault diagnosis path, and output the top-ranked fault diagnosis path as the diagnosis scheme.

[0131] Preferably, this step can include:

[0132] The fault diagnosis query information is matched with the supergraph nodes of the multi-scale fusion supergraph based on the multi-scale fusion supergraph to generate a candidate diagnosis path. Optionally, the input fault diagnosis query information is semantically parsed, key semantic features are extracted, and vectorized representations are obtained, the semantic similarity between the vectorized representations of the query information and each node in the multi-scale fusion supergraph is calculated, and nodes with a similarity higher than a preset threshold are selected as starting nodes of the candidate path; starting from the starting nodes, the path is expanded according to the superedge connection relationship to generate a plurality of possible candidate diagnosis paths, each path including a plurality of nodes and connection relationships, forming a complete diagnosis reasoning chain.

[0133] The structure consistency and semantic rationality of each candidate diagnosis path are evaluated, and the candidate diagnosis paths are sorted according to the evaluation results. Optionally, the structure topology of each candidate diagnosis path is analyzed to determine whether the connection between the nodes and the superedges conforms to the preset structure constraint rule, ensuring the structural rationality of the path; the semantic coherence and logical rationality of each candidate path are evaluated in combination with the node semantic features and the path context information to avoid semantic conflicts or logical breaks; according to the evaluation results of the structure consistency and the semantic rationality, a comprehensive score is given to each candidate path, and the candidate paths are sorted in descending order of the score, and a plurality of paths with high rankings are selected.

[0134] The candidate diagnosis paths with high rankings are output as diagnosis schemes. Optionally, according to the sorting results, a plurality of candidate paths with the highest comprehensive scores are selected as optimal diagnosis paths, and the optimal diagnosis paths are confirmed manually or automatically, and the confirmed optimal diagnosis paths are output in a structured form. Optionally, the output diagnosis schemes are analyzed for feedback, and optimization suggestions are provided based on the actual diagnosis effect, for iterative improvement of the subsequent model and improvement of the diagnosis accuracy.

[0135] The scheme generates and sorts the fault diagnosis paths based on the multi-scale fusion supergraph, realizes efficient construction and accurate output of the diagnosis paths. The candidate paths are quickly generated by semantic matching and superedge connection relationship, and double evaluation is performed in combination with structure consistency and semantic rationality to ensure that the diagnosis scheme is both structurally reasonable and semantically coherent, significantly improving the accuracy and explainability of the diagnosis. The method supports intelligent decision-making in complex scenarios and provides reliable and understandable high-quality diagnosis schemes for users. At the same time, a closed-loop mechanism is formed through feedback and optimization suggestions to promote continuous iterative optimization of the model, and the adaptability and robustness of the system in actual application are enhanced.

[0136] The present application realizes the intelligent processing of the whole process from the natural language fault query to the accurate diagnosis scheme. First, the user query is converted into structured semantic guidance information through semantic analysis, effectively capturing the intention and constraints; second, the task path is dynamically constructed based on multi-scale nested sub-graphs, and the complex task nodes and their relationships are expressed in multiple dimensions using hypergraph modeling technology, which improves the model's representation ability of complex systems. Subsequently, through structure perception regularization and cross-scale attention mechanism, the hypergraph structure and semantic fusion are optimized to ensure the structural consistency and semantic rationality of the diagnosis path. Finally, based on the multi-scale fusion hypergraph, the fault diagnosis path is generated and sorted, and the high-quality diagnosis scheme is output. The method significantly improves the accuracy, robustness and explainability of diagnosis, is suitable for intelligent decision support in complex industrial scenarios, and has good scalability and self-optimization ability.

[0137] In another aspect, the embodiments of the present application provide a system, comprising a storage medium and one or more processors; the storage medium stores a computer program, and the computer program is called by the one or more processors to implement any of the above methods.

[0138] In an exemplary embodiment, a computer-readable storage medium is also provided, the storage medium storing at least one program code, the at least one program code being loaded and executed by a processor to enable a computer to implement any of the above methods.

[0139] Optionally, the above computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0140] In an exemplary embodiment, a computer program or computer program product is also provided, the computer program or computer program product storing at least one computer instruction, the at least one computer instruction being loaded and executed by a processor to enable a computer to implement any of the above methods.

[0141] It should be noted that all actions of obtaining signals, information or data in the present application are carried out in accordance with the data protection regulation policy of the country where the device is located and with the authorization given by the owner of the corresponding device.

[0142] It should be understood that the "multiple" mentioned herein refers to two or more than two. The "and / or" describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the front and rear associated objects are in an "or" relationship.

[0143] The above-mentioned sequence numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0144] The above-mentioned only for the exemplary embodiments of the present application, and does not limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application, should be included in the protection scope of the present application.

Claims

1. A method for generating diagnostic schemes based on multi-scale hypergraphs and constrained optimization, characterized in that, The method includes: Receive fault diagnosis query information expressed in natural language, and perform semantic analysis on the fault diagnosis query information to form semantic guidance information; Based on the semantic guidance information, task nodes are extracted from the technical information set, and a multi-scale nested subgraph is constructed according to the relationship between the task nodes. Based on multi-scale nested subgraphs, the task path is dynamically reconstructed, and the task nodes are represented as hypergraph nodes to construct a multi-scale hypergraph. The multi-scale hypergraph is subjected to structure-aware regularization to obtain a regularized multi-scale hypergraph. Cross-scale attention processing is performed on the regularized multi-scale hypergraph to obtain a multi-scale fused hypergraph; Fault diagnosis paths are generated based on multi-scale fusion hypergraphs, and the fault diagnosis paths are sorted. The top-ranked fault diagnosis paths are output as the diagnosis schemes. The process of performing structure-aware regularization on the multi-scale hypergraph to obtain a regularized multi-scale hypergraph includes: Calculate the semantic similarity score between nodes and hyperedges of the multi-scale hypergraph. For each node, select the hyperedge with the highest semantic similarity score and construct the sparse connection matrix of the multi-scale hypergraph. Based on the sparse connection matrix, superedges with connection strength greater than or equal to a preset threshold are retained to form a pruned sparse connection matrix. By constructing constraints, logical verification and conflict reconciliation are performed on the pruned sparse connection matrix to obtain the regularized multi-scale hypergraph. The step of performing logical verification and conflict reconciliation on the pruned sparse connectivity matrix by constructing constraints to obtain the regularized multi-scale hypergraph includes: A structurally constrained hypergraph is constructed, which includes constraints on the nodes and hyperedges of the multi-scale hypergraph. The structurally constrained hypergraph is modeled using a hypergraph neural network to learn node selection weights and hyperedge constraint tensions, thereby obtaining a structural consistency optimization model. Based on the structural consistency optimization model, the pruned sparse connection matrix is ​​logically verified and conflict reconciled to obtain a regularized multi-scale hypergraph. The step of performing logical verification and conflict reconciliation on the pruned sparse connectivity matrix based on the structural consistency optimization model to obtain a regularized multi-scale hypergraph includes: The structural consistency optimization model is used to perform logical verification on the pruned sparse connection matrix to identify node-to-hyperedge connection pairs that violate the constraints. The identified conflicting connection pairs are reconciled, including reallocating node selection weights, adjusting hyperedge constraint tensions, or eliminating redundant connections to eliminate logical conflicts. Update the sparse connectivity matrix based on the harmonic results and output a regularized multi-scale hypergraph.

2. The diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization according to claim 1, characterized in that, The step of performing semantic analysis on the fault diagnosis query information to form semantic guidance information includes: The fault diagnosis query information is subjected to intent recognition and semantic decomposition to obtain semantic representation information; The semantic representation information is subjected to contextual semantic modeling and entity extraction processing to obtain structured information; Reasoning is performed on the structured information to obtain potential constraint information, and the structured information and the potential constraint information form the semantic guidance information.

3. The diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization according to claim 1, characterized in that, The step of extracting task nodes from the technical information set based on the semantic guidance information and constructing a multi-scale nested subgraph according to the relationships between the task nodes includes: Based on the semantic guidance information and the technical information set, a structured matching process is performed to extract task nodes, which include core diagnostic targets, related components, and key operating condition features. Based on the parent-child relationships and dependencies between the task nodes, a multi-scale nested subgraph is constructed, wherein each of the multi-scale nested subgraphs corresponds to a task structure at a different level of abstraction.

4. The diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization according to claim 3, characterized in that, The method of dynamically reconstructing the task path based on multi-scale nested subgraphs, representing the task nodes as hypergraph nodes, and constructing a multi-scale hypergraph includes: Based on the semantic guidance information and the task node relationship in the multi-scale nested subgraph, the node sequence of the task path is dynamically adjusted to obtain the reconstructed task path; The task nodes in the multi-scale nested subgraph are mapped to hypergraph nodes, and the attribute vectors of the hypergraph nodes include semantic features, structural features, and constraint features. Based on the reconstructed task path and the hypergraph nodes, a multi-scale hypergraph is constructed.

5. The diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization according to claim 1, characterized in that, The process of constructing constraints to logically verify and reconcile conflicts in the pruned sparse connectivity matrix to obtain the regularized multi-scale hypergraph includes: Construct a structurally constrained hypergraph, which includes constraints on the nodes and hyperedges of the multi-scale hypergraph; The structural constraint hypergraph is modeled using a hypergraph neural network, and node selection weights and hyperedge constraint tensions are learned to obtain a structural consistency optimization model. Based on the structural consistency optimization model, logical verification and conflict reconciliation are performed on the pruned sparse connection matrix to obtain a regularized multi-scale hypergraph.

6. The diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization according to claim 1, characterized in that, The process of performing cross-scale attention processing on the regularized multi-scale hypergraph to obtain a multi-scale fused hypergraph includes: For each structural scale, the semantic attention weight between each node of the multi-scale hypergraph and the hyperedge connected to that node is calculated. Based on the semantic attention weight, node information is aggregated and node representation is updated to obtain the fused hypergraph nodes. Linear mapping is performed on the hyperedges of the multi-scale hypergraph at each structural scale, and the hyperedges of different structural scales are fused to obtain the fused hypergraph hyperedges. The fused hypergraph nodes are spliced ​​with the fused hypergraph hyperedges to obtain the multi-scale fused hypergraph.

7. The diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization according to claim 1, characterized in that, The process of generating fault diagnosis paths based on multi-scale fusion hypergraphs, ranking these paths, and outputting the top-ranked fault diagnosis paths as diagnostic schemes includes: Based on the multi-scale fusion hypergraph, the fault diagnosis query information is matched with the hypergraph nodes of the multi-scale fusion hypergraph to generate candidate diagnosis paths; Each candidate diagnostic path is evaluated for structural consistency and semantic rationality, and the candidate diagnostic paths are ranked according to the evaluation results. The candidate diagnostic paths that are ranked highest are output as diagnostic schemes.

8. A system comprising a storage medium and one or more processors; said storage medium storing a computer program, characterized in that, The computer program is invoked by one or more processors to implement the diagnostic scheme generation method based on multi-scale hypergraph and constraint optimization as described in any one of claims 1 to 7.

Images