A method for autonomous incremental update of causal nodes in a planetary gearbox of a wind turbine

Abstract

This invention discloses an autonomous incremental update method for causal nodes in a wind turbine planetary gearbox, comprising the following steps: Step S1: Extracting the pre-fault analysis interval, constructing the current window relationship structure, and obtaining a relationship change graph; generating candidate nodes based on the relationship change graph and performing lightweight screening to obtain the current set of candidate nodes; Step S2: Integrating the candidate nodes into the extended working graph, maintaining the long-term formal main graph and the extended working graph in parallel, and performing incremental updates at the structural layer; Step S3: Matching and merging candidate nodes into candidate prototypes, performing node promotion judgment, and simultaneously performing edge promotion judgment on candidate edges, promoting those that meet the conditions to formal nodes and formal edges. This invention can autonomously discover and incrementally absorb newly added fault causal mechanisms under long-term operation conditions with small samples and varying operating conditions, while stably maintaining historical diagnostic knowledge, achieving highly interpretable and continuously updated gearbox fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis and intelligent operation and maintenance technology for wind power generation equipment, specifically to a method for autonomous incremental update of causal nodes in a planetary gearbox of a wind turbine. Background Technology

[0002] Offshore wind turbines operate in environments characterized by high humidity, high salt spray, strong wind disturbances, and complex alternating loads, facing risks of high failure rates and significant impacts after a failure. Their critical transmission components, especially the planetary gearbox, are responsible for energy transfer and speed conversion from the low-speed rotor side to the high-speed generator side. Their complex internal structure and tight coupling mean that a failure can often lead to severe consequences such as complete turbine shutdown, increased maintenance costs, and increased power generation losses. Therefore, conducting research on highly reliable, interpretable, and sustainably updated fault diagnosis for offshore wind turbine planetary gearboxes has significant engineering value and application implications.

[0003] Obtaining real-world fault samples for offshore wind turbines is challenging, especially since the various fault modes of planetary gearboxes occur infrequently in actual operation and exhibit significant differences between batches. This makes it impractical to rely solely on training high-precision models using large-scale labeled samples. Meanwhile, practical diagnostic tasks not only prioritize classification accuracy but also emphasize interpretability and traceability. Models are expected to explain the source of anomalies, key influencing factors, and potential propagation paths to support operational decisions, alarm handling, and maintenance planning. Particularly in continuous operation scenarios, diagnostic models need to address a more fundamental issue: how can they update themselves without disrupting existing knowledge when new anomaly patterns, potential fault mechanisms, or the gradual failure of existing mechanisms emerge? Therefore, a fault diagnosis method for offshore wind turbine planetary gearboxes that is continuously updated and possesses structural interpretability under limited sample conditions has become a crucial requirement in the field of intelligent diagnostics.

[0004] Currently, for fault diagnosis of offshore wind turbine planetary gearboxes under limited fault samples, complex variable operating conditions, and high interpretability requirements, a major approach in existing technologies is mechanistic modeling methods based on variable relationship graphs or causal graphs. For example, correlation analysis and other methods are used to identify direct interactions between variables and construct a system relationship network, thereby locating potential propagation paths when anomalies occur. In recent years, some studies have also attempted to separate stable and variable relationships between different operating states, wind speed ranges, or load stages, adjusting the fault diagnosis graph structure to adapt to the state drift problem during long-term operation of offshore wind turbines. However, existing technologies typically only assess the relationships between pre-defined object nodes in the "relationship change detection" stage, lacking consideration for new objects generated by changes in operating conditions, thus hindering the algorithm's ability to autonomously adapt to new environments.

[0005] To address the challenge of continuous model adaptation to changing data distributions, academia and engineering have proposed various incremental and continuous learning methods. For example, parameter regularization constrains the weights of historical knowledge, or sample replay and memory buffer mechanisms maintain performance from previous tasks, thus mitigating catastrophic forgetting. Furthermore, incremental modeling schemes based on dynamic network expansion or task branching structures enable the model to add new representation units to adapt to changing distributions when new data arrives. These methods have achieved some success in image recognition and prediction modeling tasks, but they primarily focus on continuous updates at the model parameter level, offering limited support for the evolution of structured knowledge, particularly causal relationships. For the complex variable operating conditions of offshore wind turbine planetary gearboxes, existing research struggles to implement reliable structural incremental learning schemes.

[0006] In summary, most existing incremental learning methods for offshore wind turbine gearboxes revolve around "how the model maintains the performance of old tasks after learning new tasks," focusing more on continuous training rather than the incremental update problem of the structural layer for causal node lifecycle management. When new abnormal correlation patterns emerge during the long-term operation of offshore wind turbines, relying solely on parameter constraints or sample replay is insufficient to stably maintain an interpretable set of causal nodes. Existing causal diagnosis or causal graph learning methods often require repeatedly learning the entire causal graph, resulting in high computational costs. Furthermore, existing methods lack a complete node-level management mechanism, limiting the learning objects to relationships between fixed objects and neglecting to consider different objects appearing under different operating conditions, thus restricting the adaptability of fault diagnosis algorithms to real-world environments. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention aims to provide an autonomous incremental update method for causal nodes of planetary gearboxes of wind turbines. This method can autonomously discover and incrementally absorb new fault causal mechanisms under long-term operation conditions with small samples and varying operating conditions, while stably maintaining historical diagnostic knowledge, thereby achieving highly interpretable and continuously updated gearbox fault diagnosis.

[0008] To achieve the objective of this invention, the following solution is adopted: A method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox includes the following steps: Step S1: During the operation of the planetary gearbox of the wind turbine, when a fault is detected, a pre-fault analysis interval is extracted before the trigger point. Based on the multivariate data window within the pre-fault analysis interval, a current window relationship structure is constructed and compared with the normal reference relationship diagram under the corresponding operating condition to obtain a relationship change diagram. Based on the relationship change diagram, at least one candidate node is generated by extracting significant change objects and aggregating common activation regions. Lightweight screening is performed on the candidate node to obtain the current round candidate node set. Step S2: Obtain the pre-maintained long-term formal main graph; connect the candidate nodes in the current round candidate node set to an extended working graph, and maintain the long-term formal main graph and the extended working graph as a parallel two-layer graph structure, performing incremental updates of the structure layer; wherein, the long-term formal main graph is used to output the causal graph, and the extended working graph is used for local structure exploration; after the update is completed, extract the local structure information of each candidate node in the extended working graph and the candidate edges in the extended working graph; Step S3: Based on the local structure information, match and merge the candidate nodes in the current round candidate node set into candidate prototypes, and perform node promotion judgment on the merged candidate prototypes. At the same time, perform edge promotion judgment on the candidate edges, promote the candidate prototypes that meet the promotion conditions to formal nodes, and promote the candidate edges that meet the promotion conditions to formal edges, and write them into the next round of formal node set and formal edge set; and perform old weak node processing on existing formal nodes to complete the incremental update of causal nodes in this round.

[0009] Further, generating at least one candidate node in step S1 specifically includes: defining each candidate node as a quadruple containing the pre-fault activation trajectory, the set of source variables, the set of source edge changes, and necessary meta-information; The pre-fault activation trajectory is used to describe the intensity change process of the candidate node in the pre-fault stage. The source variable set is used to represent the variable cluster that supports the generation of the candidate node. The source edge change set is used to represent the relation change region that nominated the candidate node. The necessary metadata includes at least the working condition label, source window number, event number, timestamp, and current prototype label.

[0010] Further, step S3, which involves matching and merging candidate nodes in the current round of candidate node set into candidate prototypes, specifically includes: Calculate the matching score between the candidate nodes in the current round of candidate node set and the existing candidate prototypes; The matching score integrates similarity of source variable set, similarity of source edge change set, similarity of pre-fault activation trajectory, similarity of local connection pattern, similarity of structural role, and similarity of edge direction set. When the matching score exceeds a preset threshold, the candidate node is merged into the corresponding candidate prototype, and the occurrence frequency, working condition distribution, and structural template of the prototype are updated; when the matching score does not exceed the preset threshold, a new candidate prototype is created and the candidate node is included in it.

[0011] Furthermore, the lightweight screening in step S1 specifically includes: Calculate the redundancy between the candidate node and the existing formal node or historical candidate prototype, calculate the precursor persistence index of the candidate node in the pre-failure stage, and calculate the condition-induced bias ratio that characterizes the bias between the operating condition label and the fault interpretation capability. The candidate node is retained in the current round of candidate node set only when the redundancy is below the first threshold, the precursor persistence index is above the second threshold, and the operating condition induced bias ratio is below the third threshold.

[0012] Furthermore, step S2 involves incremental updates of the structural layer, specifically including: Extract the basic representation of the current round node and combine it with the reference representation of the historical formal nodes to construct stable branch representations and variable branch representations; The adjacency matrix of the long-term formal main graph is constructed primarily using the stable branch representation and secondarily using the variable branch representation; the adjacency matrix of the extended working graph is constructed primarily using the variable branch representation and secondarily using the stable branch representation, and the structural updates are restricted to local regions that correspond to the candidate node source variable set by allowing an update mask matrix.

[0013] Furthermore, step S3 involves performing a node promotion determination on the merged candidate prototypes, specifically including: Define a node promotion score for a candidate prototype. The node promotion score comprehensively represents the cross-event reproducibility of the candidate prototype in multiple events, and the local structural consistency represents the degree of consistency of the local structural patterns of each candidate instance within the candidate prototype in the extended working graph. When the node's promotion score exceeds the fourth threshold, the candidate prototype is promoted to a formal node.

[0014] Furthermore, step S3 involves performing a promotion determination on the relevant candidate edges, specifically including: Define an edge promotion score for a candidate edge connecting two nodes. The edge promotion score comprehensively represents the directional stability of the edge in multiple events and the connection object consistency, which represents the stability of the set of objects connected by the edge. When the edge promotion score exceeds the fifth threshold, the candidate edge is promoted to a formal edge. Node promotion and edge promotion are performed separately.

[0015] Furthermore, the parallel maintenance of the two-layer graph structure in step S2 is specifically as follows: The node set of the long-term formal main graph contains only long-term formal nodes, and its generation logic is based primarily on stable relationships and secondarily on changing relationships; the node set of the extended working graph contains formal nodes and currently waiting nodes, and its generation logic is based primarily on changing relationships and secondarily on stable relationships. The long-term formal master graph is used to output the current causal graph to the diagnostic system, and the extended working graph is used to explore emerging structures in local regions supported by evidence.

[0016] Furthermore, step S3 involves processing existing formal nodes as old weak nodes, specifically including: Based on historical data, the activation degradation degree and contribution degradation degree of each formal node are checked, and nodes that exceed the threshold are selected into the weak old node set. The structural offset of the nodes in the weak old node set is judged, and the nodes with high structural offset are included in the observation set and structural drift relief constraints are applied. Degenerate nodes that have no obvious structural changes or are already in the observation set will be removed from the official node set.

[0017] Furthermore, the fault triggering in step S1 specifically includes: Define a joint trigger statistic, which integrates the deviation of key performance indicators and the degree of joint anomaly of multiple variables; When the joint trigger statistics exceed the sixth threshold, the current moment is determined to be the fault trigger point, and the subsequent causal node incremental update process is triggered.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention enhances the interpretability and traceability of gearbox fault mechanism characterization. By combining pre-fault relationship change extraction, candidate node construction, and a two-layer structure update mechanism of main graph and extended working graph, this invention can clearly map newly added characterizations to local structural offsets, supporting variable sets, and their evolution trajectories, avoiding reliance solely on black-box features for diagnosis. This improves the interpretability of fault propagation paths, key nodes, and sources of structural changes.

[0019] 2. This invention improves diagnostic matching capabilities in real-world environments with small sample sizes. By utilizing a pre-fault window for candidate nomination after a fault is triggered and combining it with lightweight screening, this invention can identify representative new mechanism nodes under conditions of limited fault samples and small batches of incremental arrivals. This reduces the risk of false nominations caused by occasional anomalies, scarce samples, or mixed operating condition labels, thereby more accurately matching the real-world fault diagnosis needs of wind turbine planetary gearboxes.

[0020] 3. This invention improves the adaptability to real-world changing operating conditions and the stability of structural updates. By constructing a two-layer graph structure that maintains a long-term formal main graph and an extended working graph in parallel, and by employing controlled access of candidate nodes, promotion of candidate prototypes and candidate edges, and handling mechanisms for old and weak nodes, this invention can achieve progressive updates of the causal graph under conditions of continuous evolution and gradual drift of the system structure, avoiding knowledge loss, structural oscillations, and graph bloat caused by frequent whole-graph reconstruction.

[0021] 4. This invention enhances the engineering practicality and robustness in long-term online operation scenarios. By performing old weak node processing on existing formal nodes, this invention can distinguish between "true node failure" and "temporary functional suppression caused by changes in graph structure," thereby reducing the accidental deletion of old nodes, maintaining the consistency, compactness, and diagnostic performance stability of the long-term graph structure, and is more suitable for online fault diagnosis and continuous operation and maintenance scenarios of wind turbine planetary gearboxes. Attached Figure Description

[0022] Figure 1 This is a flowchart of the autonomous incremental update method for causal nodes of a wind turbine planetary gearbox in an embodiment of the present invention; Figure 2 This is a flowchart of the candidate instance nomination and lightweight screening under fault triggering in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the incremental update of the two-layer graph structure after a candidate access in this embodiment of the invention. Figure 4 This is a flowchart of the candidate prototype promotion and old weak node handling in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0024] Terminology Explanation: 1. Causal Nodes: These are graph-structured nodes used to characterize fault evolution mechanisms, structural anomaly patterns, or local action mechanisms. These nodes do not simply correspond to the original sensor variables, but are defined by changes in variable relationships, local structural shifts, anomalous activation trajectories, and their contextual information. They serve as the basic representation units for fault interpretation, structural inference, and incremental updates in the graph model.

[0025] 2. Autonomous Incremental Learning: This refers to the system's continuous learning process, which involves autonomously completing tasks such as nominating candidate nodes, screening nodes, updating the structure, promoting nodes, observing weak nodes, and managing deletions, without relying on frequent manual reconstruction of the entire model. This process emphasizes the controlled absorption of newly added fault mechanisms while retaining valid historical knowledge.

[0026] 3. Graph-shift (Relationship Change Graph or Structural Shift Graph): This graph compares the variable relationship structure in the window before the failure with the normal reference relationship graph under the corresponding operating condition. It describes the local structural shift of the current window relative to the normal mechanism. This graph is not the final causal graph, but rather an intermediate representation used for nominating candidate mechanisms, generating candidate nodes, and supporting subsequent incremental updates of structural layers.

[0027] like Figure 1 As shown, this embodiment of the invention provides a method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox, comprising the following steps: Step S1: During the operation of the planetary gearbox of the wind turbine, when a fault is detected, a pre-fault analysis interval is extracted before the trigger point. Based on the multivariate data window within the pre-fault analysis interval, a current window relationship structure is constructed and compared with the normal reference relationship diagram under the corresponding operating condition to obtain a relationship change diagram. Based on the relationship change diagram, at least one candidate node is generated by extracting significantly changed objects and aggregating common activation regions. Lightweight screening is performed on the candidate node to obtain the current round candidate node set.

[0028] Step S2: Obtain the pre-maintained long-term formal main graph; connect the candidate nodes in the current round candidate node set to an extended working graph, and maintain the long-term formal main graph and the extended working graph as a parallel two-layer graph structure, performing incremental updates of the structure layer; wherein, the long-term formal main graph is used to output the causal graph, and the extended working graph is used for local structure exploration; after the update is completed, extract the local structure information of each candidate node in the extended working graph and the candidate edges in the extended working graph.

[0029] Step S3: Based on the local structure information, match and merge the candidate nodes in the current round candidate node set into candidate prototypes, and perform node promotion judgment on the merged candidate prototypes. At the same time, perform edge promotion judgment on the candidate edges, promote the candidate prototypes that meet the promotion conditions to formal nodes, and promote the candidate edges that meet the promotion conditions to formal edges, and write them into the next round of formal node set and formal edge set; and perform old weak node processing on existing formal nodes to complete the incremental update of causal nodes in this round.

[0030] The following is a more detailed description of the autonomous incremental update method for causal nodes of the planetary gearbox of a wind turbine according to an embodiment of the present invention.

[0031] This invention enhances the interpretability of fault mechanism characterization. By extracting nodes from the relational changes in the multivariate data window before the fault and organizing them in conjunction with node values, structural fingerprints, support sets, and meta-information, the causal graph nodes are not limited to expert experience but originate from traceable evidence of structural changes. This mechanism ensures that subsequent process nodes more accurately correspond to the actual fault evolution mechanism of offshore wind turbine planetary gearboxes.

[0032] This invention can accurately match small sample real-world environments. Addressing the reality that offshore wind turbine planetary gearbox fault samples are limited, anomaly types occur infrequently, and it is difficult to rely on large-scale, complete labeled data for repeated training, this invention ensures that the system can gradually identify and accumulate effective fault mechanisms under limited fault samples and limited new event observations, without requiring a complete reconstruction of the entire causal graph after each new fault occurs.

[0033] This invention enhances the adaptability to real-world changing operating conditions: a two-layer update structure of formal master diagram and extended working diagram, combined with a process setting for identifying old weak nodes and periodic calibration mechanisms, effectively addresses the problems of continuously changing wind speed, load, and operating status in the complex marine environment of offshore wind turbine planetary gearboxes, the continuous evolution of variable relationships with operating conditions, and the potential temporary weakening of old nodes due to structural drift. This improves the continuous adaptability, stability, and long-term robustness of the causal diagnostic structure to real-world changing operating conditions.

[0034] The autonomous incremental update method for causal nodes in wind turbine planetary gearboxes, as described in this invention, is an autonomous incremental learning method for updating causal nodes in a continuously evolving fault diagnosis scenario. This method addresses the multivariate industrial monitoring data stream of offshore wind turbine planetary gearboxes. Under conditions of limited fault samples, continuously changing operating conditions, and potentially expanding fault mechanisms, it employs a closed-loop update mechanism of "candidate node nomination—incremental update—handling of old weak nodes" to achieve autonomous discovery, controlled access, cross-event verification, long-term inheritance, and decommissioning management of causal nodes.

[0035] In this embodiment, the specific contents of the data input and basic state quantities are as follows: Before performing node updates, the raw monitoring data of the offshore wind turbine planetary gearbox is first preprocessed. Assume the system operates at discrete time intervals. The monitoring vector is: in, To monitor the dimensionality of variables.

[0036] After performing time alignment, missing value imputation, outlier suppression, and normalization on the original time series data, a length of [length missing] is used. Step size is The sliding time window is used to construct the analysis window. Each window is denoted as The corresponding data matrix is ​​denoted as If the current analysis object corresponds to the first For each subsequent fault event, its operating condition label is recorded as follows: The event number is recorded as For each operating condition, a corresponding normal reference relationship diagram is pre-constructed using normal samples, and its diagram object is denoted as... The adjacency matrix is ​​denoted as At the same time, save the previous long-term official main image. and its adjacency matrix It also maintains the historical representation cache, historical activation statistics, historical interpretation contribution statistics, and historical structural role templates of each official node, which serve as the basis for subsequent incremental updates and old node calibration.

[0037] like Figure 2 As shown in this embodiment, the specific details of the candidate instance nomination and lightweight screening under fault triggering are as follows: During the operation of an offshore wind turbine planetary gearbox, when key performance indicators or the overall behavior of multiple variables deviate from the normal range, fault analysis of the current gear is triggered. To achieve a unified representation of the triggering conditions, a joint triggering statistic is defined as follows: in, Indicates the degree of deviation of key performance indicators. Indicates the degree of joint abnormality of multiple variables. For weighting coefficients. When At that time, the determination time This is the trigger point for this round of fault analysis.

[0038] After the trigger point is determined, a pre-fault analysis interval is extracted before the trigger point. The purpose of this setting is to extract mechanistic clues from structural changes prior to the fault formation as much as possible, avoiding directly writing post-fault anomalies into the causal graph. For any window within the pre-fault interval... Correlation analysis is used to estimate the variable relationship matrix within the current window. and the normal reference relationship matrix under the corresponding working conditions. Comparison yields a graph of the changing data. Used to characterize the local structural offset of the current window relative to normal operating conditions.

[0039] Threshold filtering is performed on objects showing significant changes in the relationship transformation graph, and co-activated regions are further aggregated to form candidate instances. For any candidate instance... Its data structure is defined as follows: in, The pre-fault activation trajectory is used to describe the intensity change process of the candidate instance during the pre-fault phase. This is the set of source variables, used to indicate which variable clusters support the generation of this candidate instance; This is the set of source edge changes, used to indicate which relational change regions support the nomination of the candidate instance; Required metadata includes at least the operating condition label, source window number, event number, timestamp, and current prototype label. Based on and Derived coarse-grained presignature It is used for fast comparison at the instance level.

[0040] After a candidate instance is generated, it is not directly written to the official main graph. Instead, a lightweight screening is performed to determine whether the instance is eligible to enter the extended working graph. Let the redundancy between the candidate instance and existing official nodes or historical candidate prototypes be: in, This is the set of official nodes from the previous round. The similarity is calculated by combining the set of source variables, the set of source edge changes, the activation trajectory, and the pre-signature. This value is used to determine whether the current candidate instance highly overlaps with the existing formal node structure.

[0041] To ensure that the candidate instance truly represents an observable local mechanism during the pre-failure phase, a precursor persistence metric is further defined: in, This represents the number of pre-fault windows. This is the activation threshold. This metric is used to constrain candidate instances to have a non-transient existence during the pre-fault phase.

[0042] To suppress the interference of work condition identity information on candidate instance nomination, the work condition induced bias ratio is redefined: in, Indicates mutual information, For operating condition labels, For fault labels, This is a very small constant. If a candidate instance is more sensitive to operating condition labels but less capable of interpreting faults, then its... It will be too large. The larger value indicates that the candidate instance may originate from changes in operating conditions and should not be included in the candidate node set.

[0043] Based on redundancy, predecessor persistence, and operating condition-induced bias, a screening process is performed on candidate instances. When When both are established, alternate instances are retained. This constitutes the current set of candidate nodes: like Figure 3 As shown in this embodiment, the specific content of the incremental update of the two-layer graph structure after the candidate access is as follows: After passing a lightweight screening, candidate instances are not directly written to the long-term formal layer, but are first integrated into the extended working graph. The current round of long-term formal main graph is denoted as... Its adjacency matrix is ​​expressed as Its node set is The extended working diagram is denoted as... Its adjacency matrix is ​​denoted as Its node set is Therefore, the main graph and the extended working graph constitute a two-layer graph structure maintained in parallel. The main graph is responsible for the inheritance and output of the long-term formal structure, while the working graph is responsible for the exploration of the local structure and the extraction of evidence with the participation of the current candidate nodes.

[0044] In the structure learning phase, the basic representation of the current round node is first extracted. And combined with historical official node references Constructing stable branch representations With variation branch representation : in, Used for extraction and Structural information that is relatively stable across states; Used to extract the current state Compared to New or significantly changed structural information.

[0045] The main graph assumes the responsibility of the long-term formal structure layer; therefore, its generation logic primarily uses stable branches and secondarily uses variable branches. The main graph update for the formal-formal part is written as follows: in, To generate operators for stable relations, To generate operators for changing relations, For the change branch correction coefficient, Formal-formal subgraph mask. This is a directed acyclic constraint projection operator. The main graph is primarily derived from stable relations, while changing relations are only included as local correction terms.

[0046] The extended working graph is used to extract local and structural evidence in the current round, so its generation logic is mainly based on changing branches and secondarily on stable relationships.

[0047] by Based on this framework, the extended working graph allows updates to the following objects: formal nodes that have a direct mapping, variable overlap, or one-hop neighborhood correspondence with the candidate node source variable set; other candidate nodes adjacent to the candidate node source region, or nominated by regions with similar relationship changes in the same event; and existing edges within the formal node subgraph and their local neighborhoods. This constraint limits the current round of structural exploration to locally supported regions. Therefore, the adjacency matrix of the extended working graph is written as: in, , , These correspond to edge blocks from official to candidate, candidate to official, and candidate to candidate, respectively. All of these edge blocks are uniformly generated by the variation branch under mask constraints. in, To generate operators for changing relations, To allow updates to the mask matrix, this restricts candidate instances to generate edges only within the scope of objects that are allowed to be updated.

[0048] The two-tiered update of the main diagram and the extended working diagram ensures the conservative inheritance of the formal structural layer and the controlled exploration of the new structure.

[0049] After the expanded working graph is updated, the local structural information of each candidate node in the working graph is extracted and used as the promotion criterion for the corresponding candidate prototype. For any candidate node... Its local structural information is defined as: in, This is a local edge connection mode. As a structural role, Let be the set of edge directions. Therefore, the complete description of the candidate node is written as: This structural information is used for subsequent merging of candidate prototypes, updating of prototype structure templates, and handling of old weak nodes.

[0050] like Figure 4 As shown in this embodiment, the specific details of the candidate prototype promotion and the handling of old weak nodes are as follows: After extracting the structural information, the candidate nodes and candidate prototypes are matched and merged. Let's assume a candidate instance... With the candidate prototype The matching score is: in, These are the weighting coefficients. For similarity of source variable sets, For the similarity of the source edge change set, Similarity to the activation trajectory before the fault. For local edge pattern similarity, For structural role similarity, Let the similarity of the edge direction set be considered. When... At that time, Merge into the corresponding candidate prototype; otherwise, create a new candidate prototype. After merging, update the occurrence count, working condition distribution, structural template, edge direction statistics, and edge strength statistics of the prototype.

[0051] The promotion of candidate prototypes is determined by a node evaluation process. Promotion is not solely determined by the number of occurrences, but rather by a comprehensive assessment of cross-event reproducibility and local structural consistency. The node promotion score is defined as: in, Indicates cross-event reproducibility, used to characterize whether the candidate prototype recurs in multiple similar events; This indicates local structural consistency, used to characterize whether the local structural patterns formed by each candidate instance within the candidate prototype are consistent in the extended working graph. These are the weighting coefficients.

[0052] Cross-event reproducibility can be defined as: in, Indicates a candidate prototype The number of times it is supported by different events. This is the normalized reference number.

[0053] Local structural consistency can be defined as in, Indicates belonging to the prototype The set of candidate instances, Representation of instances Local structural template, Representing the prototype Prototype-level structural template, This represents the template difference function.

[0054] when At that time, the candidate prototype Promoted to a formal node and added to the next round of formal node set. .

[0055] After node promotion, edge promotion is performed on the candidate edges associated with that prototype. Let the candidate edges be... Connecting nodes With nodes Define its edge promotion score as: in, Indicates directional stability, which characterizes whether the direction of an edge remains consistent across multiple events and multiple updates to the working graph; It indicates the consistency of connected objects and is used to characterize whether the set of objects connected by the edge is stable; These are the weighting coefficients.

[0056] Directional stability can be written as: in, and These represent the number of times the edge appears in each of the two directions in the historical structure record.

[0057] The consistency of connection objects can be written as: in, This indicates the number of times the edge maintains the same connected object across different events. This indicates the total number of times the edge appears. When... At that time, the substitute border It is promoted to a formal edge and added to the set of formal edges in the next round. Therefore, node promotion and edge promotion are carried out separately to avoid writing unstable local edges into the main graph as a whole once the candidate prototype is officially adopted.

[0058] After completing this round of prototype promotion assessment, existing formal nodes will be processed as old and weak nodes. Based on historical data, the activation degradation degree and contribution degradation degree of each formal node will be checked. Nodes exceeding the threshold will be selected into the set of weak and old nodes.

[0059] Nodes in the weak and old node set are assessed for structural offset. Nodes with high structural offset are temporarily included in the observation set and given structural drift mitigation constraints, as their degradation may be due to structural change suppression. Degraded nodes without significant structural changes or already in the observation set are removed from the official node set. This completes the incremental update cycle of the cause-effect graph for fault diagnosis of offshore wind turbine planetary gearboxes.

[0060] The embodiments of this invention design an overall process for autonomous incremental updates of causal nodes for continuously evolving conditions. The process of updating the causal structure in fault diagnosis is divided into three consecutive stages: "candidate nomination and lightweight screening under fault triggering - incremental update of the structural layer after candidate access - handling of old weak nodes". This unifies the introduction of new nodes, retention of old nodes and exit of nodes into the same closed-loop update framework.

[0061] This invention presents a mechanism for constructing candidate nodes and promoting prototypes in a causal graph: Conceptual candidate nodes are defined, and a lightweight screening standard is established to extract candidate nodes from a multivariate data relationship change graph; conceptual candidate prototypes are defined, and a matching score formula is established so that candidate nodes with high matching scores to a particular prototype can be included in that prototype, thus advancing the prototype promotion determination; simultaneously, candidate prototype promotion criteria and candidate edge promotion criteria are established, and candidate prototypes and connected edges that meet the conditions can be promoted to formal nodes and their formal structures.

[0062] This invention proposes a two-layer structure update method that maintains the main graph and the extended working graph in parallel: based on all node representations, stable branch representations and variable branch representations are defined; the nodes of the main graph are taken from the formal node set, and are constructed with stable branches as the main component and variable branches as the auxiliary component; the nodes of the extended working graph are taken from the formal node set and the candidate node set, and are constructed based on the main graph with variable branches as the main component; the main graph is used to output the current causal graph, and the extended working graph is used to explore emerging structures.

[0063] The autonomous incremental update method for causal nodes of the planetary gearbox of a wind turbine in this embodiment of the invention has the following advantages: 1. This invention enhances the interpretability and traceability of fault mechanism characterization for offshore wind turbine gearboxes. By combining pre-fault relationship change extraction, candidate node four-element description construction, and a two-layer structure update mechanism of main graph / working graph, newly added characterizations can be clearly mapped to local structural offsets, supporting variable sets, and their evolution trajectories, avoiding reliance solely on black-box features for diagnosis. This improves the interpretability of fault propagation paths, key nodes, and sources of structural changes.

[0064] 2. The embodiments of the present invention improve the diagnostic matching capability in real-world environments with small samples. By using only the pre-fault window for candidate nomination after a fault is triggered, and combining reproducibility, persistence, redundancy, and operating condition bias gating for lightweight screening, more representative new mechanism nodes can be identified under conditions of limited fault samples and small batches of incremental arrivals. This reduces the risk of false nominations caused by sporadic anomalies, scarce samples, or mixed operating condition labels, thereby more accurately matching the real fault diagnosis needs of offshore wind turbine planetary gearboxes.

[0065] 3. The embodiments of the present invention improve the adaptability to real-world changing operating conditions and the stability of structural updates. By decoupling the modeling of stable and changing relationships, and employing controlled access of candidate nodes, separation and promotion of nodes and edges, and observation and retention mechanisms for old weak nodes, the causal graph can be progressively updated even as the operating conditions of offshore wind turbine planetary gearboxes continue to evolve and the system structure gradually drifts. This avoids knowledge loss, structural oscillations, and graph bloat caused by frequent whole-graph reconstruction.

[0066] 4. The embodiments of this invention enhance the engineering practicality and robustness in long-term online operation scenarios. By introducing structural drift judgment and periodic global calibration mechanisms for old weak nodes, it is possible to distinguish between "true node failure" and "temporary functional suppression caused by changes in graph structure," thereby reducing the accidental deletion of old nodes, maintaining the consistency, compactness, and diagnostic performance stability of the long-term graph structure, and making it more suitable for online fault diagnosis and continuous operation and maintenance scenarios of offshore wind turbine planetary gearboxes.

[0067] Compared to existing technologies, while existing techniques, such as incremental causal graph learning, can achieve online updates of causal structures through trigger point detection and reduce the cost of retraining the entire graph to some extent, their updates typically focus on adjusting relationships on existing node sets. They lack a detailed mechanism for "how to discover, screen, and controllably integrate new fault mechanisms in offshore wind turbine planetary gearboxes into the graph structure." Furthermore, general continuous learning methods primarily focus on parameter preservation or representation preservation, easily misinterpreting operating condition differences, task-induced biases, or occasional anomalies as valid knowledge, resulting in insufficient interpretative granularity of fault mechanism representation. In contrast, this technical solution extracts the relationship change graph from the pre-fault window and organizes it into candidate nodes with node values, structural fingerprints, support sets, and meta-information. Combined with redundancy, persistence, and operating condition-induced gating for lightweight screening, this ensures that newly added structural units possess clear mechanistic semantics and screening criteria before entering the formal causal graph. Therefore, it has a greater advantage in the interpretability and traceability of fault mechanism representation in offshore wind turbine planetary gearboxes.

[0068] Furthermore, existing technologies are still insufficient for coordinating the arrival of small sample increments, continuous evolution of graph structures, and stable retention of old knowledge in real-world scenarios involving offshore wind turbine planetary gearboxes. This proposed solution constructs a two-layer structure that maintains the main graph and the extended working graph in parallel. Only candidate nodes are allowed to be explored in a controlled manner within the working graph, and a node and edge separation promotion mechanism is employed to prevent main graph expansion. Simultaneously, by combining observation and deletion of old weak nodes to form a complete node update process, this fault diagnosis solution maintains better adaptability, structural stability, and long-term online operational practicality under continuously changing operating conditions and gradual structural drift.

[0069] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for autonomous incremental update of causal nodes in a planetary gearbox of a wind turbine, characterized in that, Includes the following steps: Step S1: During the operation of the planetary gearbox of the wind turbine, when a fault is detected, a pre-fault analysis interval is extracted before the trigger point. Based on the multivariate data window within the pre-fault analysis interval, a current window relationship structure is constructed and compared with the normal reference relationship diagram under the corresponding operating condition to obtain a relationship change diagram. Based on the relationship change diagram, at least one candidate node is generated by extracting significant change objects and aggregating common activation regions. Lightweight screening is performed on the candidate node to obtain the current round candidate node set. Step S2: Obtain the pre-maintained long-term official main graph; The candidate nodes in the current set of candidate nodes are connected to an extended working graph, and the long-term formal main graph and the extended working graph are maintained in parallel as a two-layer graph structure for incremental updates of the structure layer; wherein, the long-term formal main graph is used to output the causal graph, and the extended working graph is used for local structure exploration; after the update is completed, the local structure information of each candidate node in the extended working graph and the candidate edges in the extended working graph are extracted. Step S3: Based on the local structure information, match and merge the candidate nodes in the current round candidate node set into candidate prototypes, and perform node promotion judgment on the merged candidate prototypes. At the same time, perform edge promotion judgment on the candidate edges, promote the candidate prototypes that meet the promotion conditions to formal nodes, and promote the candidate edges that meet the promotion conditions to formal edges, and write them into the next round of formal node set and formal edge set; and perform old weak node processing on existing formal nodes to complete the incremental update of causal nodes in this round.

2. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, In step S1, generating at least one candidate node specifically includes: defining each candidate node as a quadruple containing the pre-fault activation trajectory, the set of source variables, the set of source edge changes, and necessary meta-information; The pre-fault activation trajectory is used to describe the intensity change process of the candidate node in the pre-fault stage. The source variable set is used to represent the variable cluster that supports the generation of the candidate node. The source edge change set is used to represent the relation change region that nominated the candidate node. The necessary metadata includes at least the working condition label, source window number, event number, timestamp, and current prototype label.

3. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, Step S3, which involves matching and merging candidate nodes in the current round of candidate node set into candidate prototypes, specifically includes: Calculate the matching score between the candidate nodes in the current round of candidate node set and the existing candidate prototypes; The matching score integrates similarity of source variable set, similarity of source edge change set, similarity of pre-fault activation trajectory, similarity of local connection pattern, similarity of structural role, and similarity of edge direction set. When the matching score exceeds a preset threshold, the candidate node is merged into the corresponding candidate prototype, and the occurrence frequency, working condition distribution, and structural template of the prototype are updated; when the matching score does not exceed the preset threshold, a new candidate prototype is created and the candidate node is included in it.

4. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, The lightweight screening in step S1 specifically includes: Calculate the redundancy between the candidate node and the existing formal node or historical candidate prototype, calculate the precursor persistence index of the candidate node in the pre-failure stage, and calculate the condition-induced bias ratio that characterizes the bias between the operating condition label and the fault interpretation capability. The candidate node is retained in the current round of candidate node set only when the redundancy is below the first threshold, the precursor persistence index is above the second threshold, and the operating condition induced bias ratio is below the third threshold.

5. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, Step S2 involves incremental updates to the structural layer, specifically including: Extract the basic representation of the current round node and combine it with the reference representation of the historical formal nodes to construct stable branch representations and variable branch representations; The adjacency matrix of the long-term formal main graph is constructed primarily using the stable branch representation and secondarily using the variable branch representation; the adjacency matrix of the extended working graph is constructed primarily using the variable branch representation and secondarily using the stable branch representation, and the structural updates are restricted to local regions that correspond to the candidate node source variable set by allowing an update mask matrix.

6. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, Step S3 involves performing node promotion determination on the merged candidate prototypes, specifically including: Define a node promotion score for a candidate prototype. The node promotion score comprehensively represents the cross-event reproducibility of the candidate prototype in multiple events, and the local structural consistency represents the degree of consistency of the local structural patterns of each candidate instance within the candidate prototype in the extended working graph. When the node's promotion score exceeds the fourth threshold, the candidate prototype is promoted to a formal node.

7. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 6, characterized in that, Step S3 involves performing edge promotion determination on the relevant candidate edges, specifically including: Define an edge promotion score for a candidate edge connecting two nodes. The edge promotion score comprehensively represents the directional stability of the edge in multiple events and the connection object consistency, which represents the stability of the set of objects connected by the edge. When the edge promotion score exceeds the fifth threshold, the candidate edge is promoted to a formal edge. Node promotion and edge promotion are performed separately.

8. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, The parallel maintenance of the two-layer graph structure in step S2 is specifically as follows: The node set of the long-term formal main graph contains only long-term formal nodes, and its generation logic is based primarily on stable relationships and secondarily on changing relationships; the node set of the extended working graph contains formal nodes and currently waiting nodes, and its generation logic is based primarily on changing relationships and secondarily on stable relationships. The long-term formal master graph is used to output the current causal graph to the diagnostic system, and the extended working graph is used to explore emerging structures in local regions supported by evidence.

9. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, Step S3 involves handling existing weak nodes, specifically including: Based on historical data, the activation degradation degree and contribution degradation degree of each formal node are checked, and nodes that exceed the threshold are selected into the weak old node set. The structural offset of the nodes in the weak old node set is judged, and the nodes with high structural offset are included in the observation set and structural drift relief constraints are applied. Degenerate nodes that have no obvious structural changes or are already in the observation set will be removed from the official node set.

10. The method for autonomous incremental update of causal nodes in a wind turbine planetary gearbox according to claim 1, characterized in that, The fault triggering in step S1 specifically includes: Define a joint trigger statistic, which integrates the deviation of key performance indicators and the degree of joint anomaly of multiple variables; When the joint trigger statistics exceed the sixth threshold, the current moment is determined to be the fault trigger point, and the subsequent causal node incremental update process is triggered.

Images