A method for predicting offshore wind turbine fault interval time based on a knowledge graph

Abstract

The application discloses a kind of offshore wind turbine fault interval time prediction methods based on knowledge graph, comprising the following steps: step one: collection operation maintenance data generates sample set;Step two: time alignment and field standardization generate structured record;Step three: construct knowledge graph to obtain node set, edge set, edge weight;Step four: establish the mapping relationship of to-be-sampled parameter vector, generate graph atlas correlation matrix and construct structured proposal distribution;Step five: generate candidate parameter vector;Step six: divide subgraph data block and establish subgraph log likelihood cache, calculate log likelihood increment;Step seven: delay accept Metropolis-Hastings iteration update to-be-sampled parameter vector to obtain posterior sample set;Step eight: output fault risk parameter and generate fault interval time prediction result.The application improves the stability and calculation efficiency of fault interval time prediction result.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power operation and maintenance and intelligent prediction technology, and in particular to a method for predicting the fault interval time of offshore wind turbines based on knowledge graphs. Background Technology

[0002] Offshore wind turbines operate under conditions of high salt spray, high humidity, and strong winds and waves. Their operating status is affected by component wear, load fluctuations, maintenance actions, and alarm events. The operation and maintenance system has accumulated a large amount of operation and maintenance data. Existing fault prediction technologies often use alarm sequences, operating condition variables, and maintenance records as inputs, and output fault probabilities, remaining lifetime estimates, or fault occurrence times through threshold rules, statistical regression, time series models, or supervised learning models. In the technical approach for predicting fault interval time, survival analysis, parameterized lifetime distribution fitting, and Bayesian inference are also used to characterize time-to-event problems to adapt to engineering conditions with limited sample size, sparse fault occurrence, and high noise.

[0003] Existing technologies often use table structures to directly splice component information, fault mode information, alarm information, and operating condition variables at the data representation level, embedding relational information in feature engineering or model weights. This makes it difficult to uniformly characterize the association structure between components, fault modes, events, and maintenance actions, resulting in insufficient statistical sharing across components and modes. The correlation between parameters is simplified to independence assumptions or low-order correlation assumptions. Knowledge graph modeling for complex systems has been used to express entity relationships and event chains. However, in fault interval prediction scenarios, knowledge graphs often remain at the level of graph construction and graph representation learning. The graph structure has not been further incorporated into the probabilistic inference process. The strength of graph relationships is difficult to transform into parameter-related constraints and proposed distribution structures, resulting in slow chain mixing, unstable convergence, and insufficient sample validity during sampling and inference.

[0004] Incomplete interval time observations are common in offshore wind power operation and maintenance data. Changes in shutdown windows, maintenance windows, and sampling intervals can lead to interval observations and right-censored observations. Existing methods often handle this by discarding samples, approximating interval information with single points, or weakening censoring constraints in likelihood construction, which increases the estimation bias of interval time distribution. When using MCMC-type methods for Bayesian inference, the general random walk proposal distribution is difficult to adapt to high-dimensional strongly correlated parameter spaces, the full likelihood calculation cost for each iteration is high, and there is a lack of graph-oriented subgraph partitioning, log-likelihood caching, and incremental update mechanisms. There is also a lack of a delayed acceptance process that combines low-cost pre-screening and high-precision acceptance judgment, making it difficult to meet the computational efficiency requirements of engineering deployment while ensuring the correctness of inference.

[0005] Therefore, how to provide a knowledge graph-based method for predicting the fault interval time of offshore wind turbines is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a knowledge graph-based method for predicting the fault interval time of offshore wind turbines. This invention constructs a knowledge graph driven by operation and maintenance data, introduces the correlation between components, fault modes, events, and maintenance actions into the probabilistic inference process, and performs parameter sampling and interval time prediction under graph structure constraints to achieve probabilistic modeling and prediction of fault interval time. This invention can effectively characterize the fault evolution law under complex correlations and censored observation conditions, and has the advantages of high prediction stability, strong consistency of parameter estimation, and controllable computational efficiency.

[0007] A knowledge graph-based method for predicting the fault interval time of offshore wind turbines according to an embodiment of the present invention includes the following steps:

[0008] Step 1: Collect operation and maintenance data of offshore wind turbines and generate a sample set;

[0009] Step 2: Perform time alignment and field standardization on the sample set to generate structured records. The structured records include component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation ranges.

[0010] Step 3: Construct a knowledge graph based on structured records to obtain the node set, edge set, and edge weights;

[0011] Step 4: Establish the parameter vector to be sampled, establish the mapping relationship between the parameter vector to be sampled and the node set, generate the graph correlation matrix based on the edge weights, and construct the structured proposal distribution based on the graph correlation matrix;

[0012] Step 5: Establish the feasible region of parameters based on the observation interval, generate truncated proposal samples within the feasible region of parameters, and obtain the candidate parameter vector;

[0013] Step 6: Divide the sample set according to the knowledge graph to obtain subgraph data blocks, establish a subgraph log-likelihood cache, and calculate the log-likelihood increment based on the candidate parameter vector;

[0014] Step 7: Using the delayed acceptance Metropolis–Hastings algorithm, based on the structured proposal distribution, subgraph log-likelihood buffer, and log-likelihood incremental iterative update of the parameter vector to be sampled, the posterior sample set is obtained when the convergence condition is met.

[0015] Step 8: Output fault risk parameters based on the posterior sample set and generate fault interval time prediction results.

[0016] Optionally, step two specifically involves:

[0017] Time alignment is performed on the sample set. The time alignment uses a uniform time unit and a preset sampling interval to map each record in the sample set to an aligned timestamp.

[0018] Perform field standardization on the sample set. Field standardization includes unifying field naming and field format.

[0019] Extract component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation ranges from the sample set to form a structured record;

[0020] Component identifiers, fault mode identifiers, event identifiers, and maintenance action identifiers are uniformly coded, and a one-to-one correspondence is established between the coded values ​​and the field values, while maintaining consistency in the sample set;

[0021] The unit of operation condition variables is standardized and missing, and the values ​​of the operation condition variables are generated and written into structured records.

[0022] The structured records are sorted based on the aligned timestamps. Fault event record pairs are determined according to the component identifier and event identifier. The time difference between the fault event record pairs is calculated, and the lower and upper bounds of the interval observation interval are determined and written into the structured records.

[0023] Optionally, step three specifically includes:

[0024] Based on the structured records, component identifiers, fault mode identifiers, event identifiers, and maintenance action identifiers are extracted. Based on the extraction results, component nodes, fault mode nodes, event nodes, and maintenance action nodes are generated and aggregated to form a node set.

[0025] The edge set is generated based on the structured record. The edge set includes the associated edges from component nodes to fault mode nodes, the associated edges from fault mode nodes to event nodes, and the associated edges from event nodes to maintenance action nodes. Each associated edge records the starting node identifier, the ending node identifier, and the edge type.

[0026] Operating condition nodes are generated based on structured records. The operating condition nodes are determined by the field names of operating condition variables and the values ​​of the operating condition variables are recorded. The associated edges from event nodes to operating condition nodes are generated and written into the edge set.

[0027] Based on structured records, interval time nodes are generated. Each interval time node records the lower and upper bounds of the observation interval. The associated edges from event nodes to interval time nodes are generated and written into the edge set.

[0028] Edge weights are calculated based on the edge set. The edge weights are determined according to the edge type and the frequency of occurrence of the corresponding field in the structured record. The edge weights are written into the edge set.

[0029] A knowledge graph is constructed based on the node set, edge set, and edge weight, and the node set, edge set, and edge weight are output.

[0030] Optionally, step four specifically includes:

[0031] A parameter vector to be sampled is established. The parameter vector is formed by arranging the parameter components in a fixed order. The number and order of the parameter components remain unchanged during the iteration process.

[0032] Establish a mapping relationship between the vector of parameters to be sampled and the set of nodes, where each parameter component is assigned a node identifier in the set of nodes.

[0033] A node relevance table is generated based on the edge set and edge weights. The node relevance table uses each pair of node identifiers in the node set as an index. The node relevance is the sum of the edge weights of the matching node identifiers of the starting node identifier and the ending node identifier in the edge set. The node relevance is zero when the edge set does not contain any matching node identifier pairs.

[0034] Based on the mapping relationship, the node correlation table is projected onto the parameter component pairs to generate the graph correlation matrix. The row index and column index of the graph correlation matrix correspond to the order of the parameter components of the parameter vector to be sampled. The elements of the graph correlation matrix take the node correlation of the parameter component pairs mapped to the node identifier pairs.

[0035] A structured proposal distribution is constructed based on the graph correlation matrix. The structured proposal distribution adopts a multivariate normal distribution. The mean vector of the structured proposal distribution is taken as the current value of the parameter vector to be sampled. The covariance matrix of the structured proposal distribution is determined by the graph correlation matrix and the step size coefficient.

[0036] Optionally, step five specifically includes:

[0037] The time interval observation interval is read from the structured record. The time interval observation interval is represented by a lower bound value and an upper bound value. The upper bound value can be a finite value or an infinite upper bound.

[0038] Establish a calculation rule for the interval time distribution corresponding to the parameter vector to be sampled. The interval time distribution calculation rule is used to calculate the interval probability that the interval time falls into the interval time observation interval under the given parameter vector to be sampled.

[0039] The interval probability is calculated one by one based on the structured records. The interval probability is the probability that the interval time is not less than the lower bound and not greater than the upper bound. Under the condition that the upper bound is infinite, the interval probability is the probability that the interval time is not less than the lower bound.

[0040] The feasible region of parameters is established based on the interval probability. The feasible region of parameters includes the values ​​of the sampled parameter vector where all interval probabilities are positive.

[0041] Proposal samples are generated based on the structured proposal distribution. Proposal samples that fall within the parameter feasible region are identified as truncated proposal samples, and these truncated proposal samples are identified as candidate parameter vectors.

[0042] Optionally, step six specifically includes:

[0043] Based on the connectivity of node sets in the knowledge graph, a subgraph identifier is determined for each structured record. The subgraph identifier is determined by the node identifiers of the component identifier, fault mode identifier, event identifier, and maintenance action identifier contained in the structured record in the node set.

[0044] Structured records are grouped according to subgraph identifiers to form subgraph data blocks. Each subgraph data block contains a subgraph identifier and a set of structured records.

[0045] Establish a subgraph log-likelihood cache. The subgraph log-likelihood cache records the log-likelihood value of the subgraph data block under the current value of the parameter vector to be sampled, using the subgraph identifier as the index.

[0046] The set of parameter change components is determined based on the candidate parameter vector and the parameter vector to be sampled. The set of parameter change components is determined by the parameter components whose values ​​are different from those of the parameter components in the candidate parameter vector and the parameter components in the parameter vector to be sampled.

[0047] Based on the mapping relationship between the parameter vector to be sampled and the node set, the parameter change component set is mapped to the change node identifier set, and the affected subgraph identifier set is determined based on the change node identifier set;

[0048] For each subgraph identifier in the affected subgraph identifier set, calculate the log-likelihood value of the subgraph data block under the conditions of the candidate parameter vector values, read the log-likelihood value corresponding to the subgraph identifier in the subgraph log-likelihood cache, and use the difference to calculate the log-likelihood increment.

[0049] Optionally, step seven specifically includes:

[0050] Read the current value of the parameter vector to be sampled, the structured proposal distribution, and the parameter feasible region. Generate proposal samples based on the structured proposal distribution. When a proposal sample satisfies the parameter feasible region condition, it is determined as a candidate parameter vector.

[0051] The proposal density ratio is calculated based on the structured proposal distribution. The proposal density ratio is obtained by dividing the proposal density of the candidate parameter vector under the condition of the parameter vector to be sampled by the proposal density of the parameter vector to be sampled under the condition of the candidate parameter vector.

[0052] Calculate the subgraph weight corresponding to the subgraph identifier in the affected subgraph identifier set based on the edge set and edge weight, set a subgraph weight threshold, and obtain a pre-filtered subgraph identifier set from the affected subgraph identifier set based on the subgraph weight threshold;

[0053] The Metropolis–Hastings algorithm with delayed acceptance is used to perform pre-screening acceptance decision. The pre-screening acceptance decision is based on the ratio of the correspondence likelihood increment to the proposal density of the pre-screened subgraph identifier set.

[0054] When the pre-screening acceptance criteria are met, the precise acceptance criteria are executed. The precise acceptance criteria generate the precise acceptance probability based on the ratio of the log-likelihood increment to the proposal density for the affected subgraph identifier set. When the precise acceptance criteria are met, the candidate parameter vector is written into the parameter vector to be sampled and the log-likelihood cache of the subgraph corresponding to the affected subgraph identifier set is updated.

[0055] Repeatedly execute candidate parameter vector generation, pre-screening acceptance determination, and precise acceptance determination to obtain the posterior sample set when the convergence condition is met.

[0056] Optionally, step eight specifically includes:

[0057] Read the posterior sample set, which contains the sample values ​​of the parameter vector to be sampled;

[0058] Fault risk parameters are calculated based on the posterior sample set. The fault risk parameters include the statistical values ​​of the parameter components of the sampled parameter vector, and the statistical values ​​of the parameter components include the mean, variance, and quantile.

[0059] Read the interval time distribution calculation rule, input the sample values ​​of the parameter vector to be sampled in the posterior sample set into the interval time distribution calculation rule, and generate the fault interval time prediction distribution;

[0060] The fault interval time prediction results are calculated based on the fault interval time prediction distribution. The fault interval time prediction results include the prediction mean, prediction median, lower bound of the prediction interval, and upper bound of the prediction interval.

[0061] Output the failure risk parameters and failure interval prediction results.

[0062] The beneficial effects of this invention are:

[0063] This invention constructs a knowledge graph using structured records and generates a graph correlation matrix and a structured proposal distribution under the constraint of the mapping relationship between the parameter vector to be sampled and the node set. This ensures that the parameter update direction is consistent with the relationship strength of the knowledge graph, thereby simultaneously characterizing the association and dependency relationships between component identifiers, fault mode identifiers, event identifiers, operating condition variables, and maintenance action identifiers within the same modeling framework. A feasible region for parameters is established based on the interval observation interval, and truncated proposal samples are generated. This ensures that the sampling process is consistent with the interval observation interval, reducing invalid iterations caused by conflicts between parameter values ​​and the observation interval, improving the stability and consistency of fault risk parameter estimation, and ultimately ensuring that the fault interval prediction distribution retains a usable probabilistic expression even under censored observation conditions.

[0064] This invention divides the sample set into subgraph data blocks based on the knowledge graph and establishes a subgraph log-likelihood cache. It calculates the log-likelihood increment by combining candidate parameter vectors. In the delayed acceptance Metropolis–Hastings algorithm, it introduces pre-screening and precise acceptance decisions, thereby constraining the allocation of computational resources. This ensures that the acceptance decision phase prioritizes using low-cost information for screening, and uses the log-likelihood increment for precise decision-making when necessary. Simultaneously, it updates the log-likelihood cache corresponding to the affected subgraph identifier. Therefore, while ensuring the closed-loop logic of the sampling iteration, it reduces the overhead of repeated likelihood calculations, improves the efficiency of unit iteration and the efficiency of generating effective posterior samples, and maintains the scalability of fault risk parameters and fault interval time prediction results under conditions of increasing engineering data scale. Attached Figure Description

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0066] Figure 1 This is a flowchart of a knowledge graph-based method for predicting the fault interval time of offshore wind turbines proposed in this invention.

[0067] Figure 2 This is a schematic diagram of the delayed acceptance Metropolis–Hastings sampling framework for a knowledge graph-based method for predicting the fault interval time of offshore wind turbines proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figure 1-2 A knowledge graph-based method for predicting the fault interval time of offshore wind turbines includes the following steps:

[0070] Step 1: Collect operation and maintenance data of offshore wind turbines and generate a sample set;

[0071] Step 2: Perform time alignment and field standardization on the sample set to generate structured records. The structured records include component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation ranges.

[0072] Step 3: Construct a knowledge graph based on structured records to obtain the node set, edge set, and edge weights;

[0073] Step 4: Establish the parameter vector to be sampled, establish the mapping relationship between the parameter vector to be sampled and the node set, generate the graph correlation matrix based on the edge weights, and construct the structured proposal distribution based on the graph correlation matrix;

[0074] Step 5: Establish the feasible region of parameters based on the observation interval, generate truncated proposal samples within the feasible region of parameters, and obtain the candidate parameter vector;

[0075] Step 6: Divide the sample set according to the knowledge graph to obtain subgraph data blocks, establish a subgraph log-likelihood cache, and calculate the log-likelihood increment based on the candidate parameter vector;

[0076] Step 7: Using the delayed acceptance Metropolis–Hastings algorithm, based on the structured proposal distribution, subgraph log-likelihood buffer, and log-likelihood incremental iterative update of the parameter vector to be sampled, the posterior sample set is obtained when the convergence condition is met.

[0077] Step 8: Output fault risk parameters based on the posterior sample set and generate fault interval time prediction results.

[0078] In this embodiment, step two specifically includes:

[0079] Time alignment is performed on the sample set. The time alignment uses a unified time unit and a preset sampling interval to convert the recorded time in the sample set into a unified time unit representation. An aligned timestamp sequence is generated according to the preset sampling interval. Each record time is mapped to the aligned timestamp with the smallest time difference in the aligned timestamp sequence. The mapping result is written into the aligned timestamp field in the structured record.

[0080] Field standardization is performed on the sample set. Field standardization includes unifying field naming and field format. Unifying field naming uses a field mapping table to map field names from different sources to standard field names for component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation intervals. Unifying field format converts identifier fields into string or integer formats, operating condition variables into numerical formats, and interval observation intervals into lower and upper bound value fields.

[0081] Component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation ranges are extracted from the sample set to form structured records. The structured records are generated at the record granularity and written into the aligned timestamp field. The structured records are aggregated within the sample set to form a structured record set.

[0082] Component identifiers, fault mode identifiers, event identifiers, and maintenance action identifiers are uniformly encoded. The encoding uses an encoding dictionary, which assigns a unique encoding value to each field value. The encoding dictionary remains fixed within the structured record set. Field values ​​in the structured records are replaced with encoding values ​​while retaining the field names.

[0083] Unit unification and missing value handling are performed on the operating condition variables. Unit unification uses unit conversion rules to convert the operating condition variables into a unified unit of measurement. Missing value handling uses missing mark rules to identify missing locations and uses preset filling rules to generate operating condition variable values. The operating condition variable values ​​are written into the operating condition variable field in the structured record.

[0084] The structured records are sorted based on the aligned timestamps. Fault event record pairs are determined according to the component identifier and the event identifier. Each fault event record pair consists of fault event records corresponding to adjacent aligned timestamps under the same component identifier. The time difference between the fault event record pairs is calculated to obtain the interval time observation value. The lower and upper bound values ​​of the interval time observation interval are determined based on the interval time observation value and written into the structured records. The lower bound value of the interval time observation interval is the interval time observation value, and the upper bound value of the interval time observation interval is either the interval time observation value or an infinite upper bound.

[0085] In this embodiment, step three specifically includes:

[0086] Based on the structured record, component identifier, fault mode identifier, event identifier, and maintenance action identifier are extracted. For each structured record, the values ​​of the component identifier field, fault mode identifier field, event identifier field, and maintenance action identifier field are read to generate component nodes, fault mode nodes, event nodes, and maintenance action nodes respectively. The node identifier is obtained from the field value, and the node type is obtained from the component node type, fault mode node type, event node type, and maintenance action node type. The nodes are deduplicated according to the node identifier and then aggregated to form a node set.

[0087] Based on the structured record, an edge set is generated. For each structured record, the component node identifier and the fault mode node identifier are located and the associated edge from the component node to the fault mode node is generated. The fault mode node identifier and the event node identifier are located and the associated edge from the fault mode node to the event node is generated. The event node identifier and the maintenance action node identifier are located and the associated edge from the event node to the maintenance action node is generated. Each associated edge records the starting node identifier, the ending node identifier, and the edge type. The edge type adopts component-fault mode, fault mode-event, and event-maintenance action.

[0088] Based on structured records, work condition nodes are generated. For each structured record, the work condition variable field name and work condition variable value are read. The work condition node identifier is determined by the combination of the work condition variable field name and the event identifier. The work condition node records the work condition variable field name and work condition variable value. The associated edges from event nodes to work condition nodes are generated and written into the edge set. The edge type adopts event-work condition.

[0089] Interval time nodes are generated based on structured records. For each structured record, the lower and upper bounds of the interval observation interval are read. The interval time node identifier is determined by the event identifier. The interval time node records the lower and upper bounds of the interval observation interval. An associated edge from the event node to the interval time node is generated and written into the edge set. The edge type is event-interval time.

[0090] Edge weights are calculated based on edge sets. For each edge in the edge set, the number of occurrences of the triplet of starting node identifier, ending node identifier, and edge type is counted as the edge count. For each edge type, the total edge count is counted as the total type count. The edge weight is determined by the ratio of the edge count to the total type count. The edge weight is written into the edge set and associated with the starting node identifier, ending node identifier, and edge type.

[0091] A knowledge graph is constructed based on a set of nodes, a set of edges, and edge weights. The knowledge graph stores node identifiers and node types in the set of nodes, and stores the starting node identifier, ending node identifier, edge type, and edge weight in the set of edges. The output is a set of nodes, a set of edges, and edge weights.

[0092] In this embodiment, step four specifically includes:

[0093] A parameter vector to be sampled is established. The parameter vector to be sampled is arranged in the order of parameter components to form a one-dimensional vector structure. The order of parameter components is determined by a preset order list. The preset order list remains fixed during the iteration process. The initial value is written into the parameter vector to be sampled during the initialization phase and the current value is written during the iteration phase.

[0094] Establish a mapping relationship between the parameter vector to be sampled and the node set. The mapping relationship is represented by a mapping table. Each row in the mapping table corresponds to a parameter component index and records the mapping node identifier. The mapping node identifier is taken from the node identifier in the node set. The mapping table remains fixed during the iteration process.

[0095] A node relevance table is generated based on the edge set and edge weights. The node relevance table adopts a two-dimensional index structure, with the row and column indices being the node identifiers in the node set, respectively. For any pair of node identifiers, the edge records matching the start node identifier and the end node identifier and the node identifier in the edge set are read. The edge weights corresponding to the matching edge records are extracted and summed to obtain the node relevance. The node relevance is set to zero when there are no matching edge records for the node identifiers. The node relevance is written to the corresponding index position in the node relevance table.

[0096] Based on the mapping relationship, the node correlation table is projected onto the parameter component pairs to generate the graph correlation matrix. The row and column indices of the graph correlation matrix are the parameter component indices of the parameter vector to be sampled. For any pair of parameter component indices, the mapping table is read to obtain the node identifier pair corresponding to the parameter component index. The node correlation corresponding to the node identifier pair is read from the node correlation table and written into the corresponding element position of the graph correlation matrix. The graph correlation matrix is ​​arranged in the order of parameter component indices to form a matrix structure.

[0097] A structured proposal distribution is constructed based on the graph correlation matrix. The structured proposal distribution adopts a multivariate normal distribution. The mean vector of the structured proposal distribution takes the current value of the parameter vector to be sampled. The covariance matrix of the structured proposal distribution is determined by the step size coefficient and the graph correlation matrix. The step size coefficient is a preset constant. The covariance matrix is ​​updated with the graph correlation matrix or remains fixed during the iteration process. The structured proposal distribution is used to generate candidate parameter vectors.

[0098] In this embodiment, step five specifically includes:

[0099] The interval observation interval is read from the structured record. The interval observation interval is stored in the structured record as an interval field. The interval field contains a lower bound field and an upper bound field. The lower bound field records the minimum interval value, and the upper bound field records the maximum interval value. The upper bound field represents the right endpoint of the interval when it is a finite value, and the upper bound field represents the right endpoint of the interval when it is an infinite upper bound.

[0100] Establish a calculation rule for the interval time distribution corresponding to the parameter vector to be sampled. The interval time distribution calculation rule defines the parameterized form of the interval time distribution and takes the parameter vector to be sampled as the parameter input. The interval time distribution calculation rule supports calculating the probability value that the interval time is not greater than the time threshold under any given time threshold, and supports calculating the interval probability that the interval time falls into the observation interval under given lower and upper bound values.

[0101] The interval probability is calculated one by one based on the structured record. For each structured record, the lower bound field and the upper bound field are read. The current value of the parameter vector to be sampled is read as the input of the interval time distribution calculation rule. The probability value that the interval time is not greater than the upper bound value and the probability value that the interval time is not greater than the lower bound value are calculated. The difference between the two is used to obtain the interval probability. The upper bound field is taken as the probability value that the interval time is not less than the lower bound value under the condition of infinite upper bound. The interval probability is written to the probability field of the structured record or written to an independent interval probability table.

[0102] The feasible region of parameters is established based on interval probability. The feasible region of parameters is represented by a set of inequality constraints. The set of inequality constraints constrains each structured record index with a positive interval probability. The parameter vector to be sampled is determined to fall into the feasible region of parameters if it satisfies all inequality constraints, and is determined not to fall into the feasible region of parameters if it does not satisfy any inequality constraints.

[0103] Proposal samples are generated based on the structured proposal distribution. The proposal sample generation process randomly selects a set of parameter component values ​​from the structured proposal distribution to form a proposal sample. For the proposal sample, the parameter feasible region is determined. If it falls into the parameter feasible region, the proposal sample is determined as a truncated proposal sample. If it does not fall into the parameter feasible region, the proposal sample is discarded and a new proposal sample is generated. The truncated proposal sample is written into the candidate parameter vector and the candidate parameter vector is output.

[0104] In this embodiment, step six specifically includes:

[0105] Based on the connectivity relationships of node sets in the knowledge graph, a subgraph identifier is determined for each structured record. For each structured record, the component identifier, fault mode identifier, event identifier, and maintenance action identifier are read. The corresponding node identifier is located by using the consistency relationship between the node identifiers in the node set and the identifier values ​​in the structured record. The node identifiers corresponding to the component identifier, fault mode identifier, event identifier, and maintenance action identifier are used as the starting node set. Within the scope of the knowledge graph node set and edge set, connectivity component determination is performed. The connectivity component determination outputs the connectivity component number as the subgraph identifier, and the subgraph identifier is written into the structured record.

[0106] Structured records are grouped by subgraph identifier to form subgraph data blocks. The subgraph identifier field in the structured record is read, and the subgraph identifier is used as the grouping key. Structured records with the same subgraph identifier are written into the same subgraph data block. The subgraph data block stores the subgraph identifier and the set of structured records.

[0107] A subgraph log-likelihood cache is established, which adopts a key-value structure. The key is the subgraph identifier, and the value is the log-likelihood value. For each subgraph data block, the subgraph identifier and the structured record set are read, the current value of the parameter vector to be sampled is read, and the log-likelihood item is calculated for each record in the structured record set according to the log-likelihood calculation rule. The log-likelihood value of the subgraph data block is obtained by summing the log-likelihood value and writing it into the subgraph log-likelihood cache at the corresponding position of the subgraph identifier.

[0108] The parameter change component set is determined based on the candidate parameter vector and the parameter vector to be sampled. The parameter component values ​​of the candidate parameter vector and the parameter component values ​​of the parameter vector to be sampled are traversed in the order of the parameter components of the parameter vector to be sampled. Under different parameter component values, the parameter component index is written into the parameter change component set. The parameter change component set stores the parameter component index set.

[0109] Based on the mapping relationship between the parameter vector to be sampled and the node set, the parameter change component set is mapped to the change node identifier set. For each parameter component index in the parameter change component set, the mapping relationship is read to obtain the node identifier. The node identifier is written into the change node identifier set. Based on the change node identifier set, the subgraph identifier query is performed in the knowledge graph. The subgraph identifier query uses the node identifier to subgraph identifier correspondence table to obtain the affected subgraph identifier set.

[0110] For each subgraph identifier in the affected subgraph identifier set, the subgraph data block is read and the log-likelihood value of the subgraph data block is calculated under the conditions of the candidate parameter vector values. The calculation process reads the structured record set in the subgraph data block, reads the candidate parameter vector values, calculates the log-likelihood terms one by one according to the log-likelihood calculation rules, and performs summation to obtain the candidate log-likelihood value. The log-likelihood value corresponding to the subgraph identifier in the subgraph log-likelihood cache is read as the cached log-likelihood value. The log-likelihood increment is obtained by subtracting the cached log-likelihood value from the candidate log-likelihood value. The log-likelihood increment is output to the delayed acceptance process of the Metropolis–Hastings algorithm.

[0111] In this embodiment, step seven specifically includes:

[0112] Read the current value of the parameter vector to be sampled, the structured proposal distribution, and the parameter feasible region. The current value of the parameter vector to be sampled is used as the current state input to the structured proposal distribution. The structured proposal distribution randomly generates proposal samples. Perform parameter feasible region determination on the proposal samples. If the parameter feasible region determination is satisfied, the proposal sample is determined as a candidate parameter vector. If the parameter feasible region determination is not satisfied, the proposal sample is discarded and a new proposal sample is generated.

[0113] The proposal density ratio is calculated based on the structured proposal distribution. The proposal density ratio is calculated by reading the current values ​​of the candidate parameter vector and the parameter vector to be sampled, calculating the proposal density of the candidate parameter vector under the condition of the parameter vector to be sampled, and calculating the proposal density of the parameter vector to be sampled under the condition of the candidate parameter vector. The former is divided by the latter to obtain the proposal density ratio. The proposal density ratio is used as the input of the proposal terms for the delayed acceptance of the Metropolis-Hastings algorithm.

[0114] The subgraph weights corresponding to the subgraph identifiers in the affected subgraph identifier set are calculated based on the edge set and edge weights. The subgraph weight calculation locates the edge records contained in each subgraph identifier within the knowledge graph edge set. The edge weights of the located edge records are extracted and summed to obtain the subgraph weight. A subgraph weight threshold is set and filtering is performed. The filtering process selects subgraph identifiers with subgraph weights not less than the subgraph weight threshold within the affected subgraph identifier set to form a pre-filtered subgraph identifier set.

[0115] The Metropolis–Hastings algorithm with delayed acceptance is used to perform pre-screening acceptance determination. The pre-screening acceptance determination reads the logarithmic likelihood increment and proposal density ratio corresponding to the pre-screened subgraph identifier set. According to the Metropolis–Hastings acceptance criterion, the proposal density ratio and the logarithmic likelihood increment corresponding to the pre-screened subgraph identifier set are combined to generate the pre-screening acceptance probability. A random number is generated using a preset random number generation rule and compared with the pre-screening acceptance probability to obtain the pre-screening acceptance determination result. If the pre-screening acceptance determination is satisfied, the precise acceptance determination is entered. If the pre-screening acceptance determination is not satisfied, the current value of the parameter vector to be sampled is maintained and the subgraph log-likelihood cache is maintained.

[0116] When the pre-screening acceptance criteria are met, a precise acceptance criteria are executed. The precise acceptance criteria read the log-likelihood increment and the proposal density ratio corresponding to the affected subgraph identifier set. Based on the Metropolis-Hastings acceptance criterion, the proposal density ratio and the log-likelihood increment corresponding to the affected subgraph identifier set are combined to generate a precise acceptance probability. A random number is generated using a preset random number generation rule and compared with the precise acceptance probability to obtain the precise acceptance criteria result. If the precise acceptance criteria are met, the candidate parameter vector is written to the parameter vector to be sampled and the log-likelihood cache of the subgraph corresponding to the affected subgraph identifier set is updated. If the precise acceptance criteria are not met, the current value of the parameter vector to be sampled is maintained and the subgraph log-likelihood cache is maintained.

[0117] Repeatedly execute candidate parameter vector generation, pre-screening acceptance judgment, and precise acceptance judgment. Stop iterating when the number of repetitions reaches the upper limit of the number of samplings or when the convergence condition is met. The convergence condition adopts the posterior sample set stability judgment rule. During the iteration process, the value of the parameter vector to be sampled under the condition of satisfying the precise acceptance judgment is written into the posterior sample set. After the iteration stops, the posterior sample set is output.

[0118] In this embodiment, step eight specifically includes:

[0119] Read the posterior sample set. The posterior sample set stores the sample values ​​of the parameter vector to be sampled by the sample index. During the reading process, the posterior sample set is traversed by the sample index and the sample value of each parameter vector to be sampled is obtained. The sample values ​​of the parameter vector to be sampled maintain the parameter component order consistent with the definition of the parameter vector to be sampled.

[0120] Fault risk parameters are calculated based on the posterior sample set. For each parameter component of the parameter vector to be sampled, the corresponding parameter component values ​​in the posterior sample set are collected to form a parameter component sample sequence. The parameter component sample sequence is used to calculate the mean, variance, and quantile. The mean is determined by the arithmetic mean of the parameter component sample sequence, the variance is determined by the average of the squared deviations of the parameter component sample sequence, and the quantile is obtained by sorting the parameter component sample sequence and selecting it according to the quantile position. The calculation results are written into the fault risk parameters.

[0121] The interval time distribution calculation rule is read. The interval time distribution calculation rule takes the value of the parameter vector to be sampled as input and outputs the interval time probability distribution calculation interface. The calculation interface supports calculating the probability that the interval time is not less than the time threshold and the probability that the interval time is not greater than the time threshold under a given time threshold. The calculation interface supports calculating the interval probability that the interval time falls into the time interval under a given time interval.

[0122] Input the sample values ​​of the parameter vector to be sampled in the posterior sample set into the interval time distribution calculation rule. Execute the interval time distribution calculation rule for each sample value of the parameter vector to be sampled and generate a set of interval time probability distribution results. Collect the interval time probability distribution results to form the fault interval time prediction distribution.

[0123] The fault interval prediction result is calculated based on the fault interval prediction distribution. A prediction sample sequence is generated for the fault interval prediction distribution. The prediction sample sequence is extracted from the interval probability distribution result according to the preset sampling rules. The prediction mean is determined by the arithmetic mean of the prediction sample sequence. The prediction median is determined by taking the median position after sorting the prediction sample sequence. The lower limit and upper limit of the prediction interval are determined by the prediction sample sequence according to the corresponding quantile position of the preset confidence level.

[0124] Output the fault risk parameters and the fault interval prediction results. The output includes the mean, variance, and quantile values ​​of the fault risk parameters, and the predicted mean, predicted median, lower bound of the prediction interval, and upper bound of the prediction interval prediction results.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to the operation and maintenance management scenario of a wind turbine in an offshore wind farm. The wind farm operates 48 turbines, each containing key components such as gearbox, generator, pitch system, yaw system, and hydraulic system. The operation and maintenance system continuously collects operation and maintenance data, with data sources covering operation monitoring, alarm events, maintenance work orders, and maintenance action records. This forms a data stream with component identifiers, fault mode identifiers, event identifiers, operating condition variables, and maintenance action identifiers as core fields. Two typical difficulties exist on-site: one difficulty stems from incomplete fault interval observations, where the downtime maintenance window makes it impossible to pinpoint the exact time of fault occurrence, and maintenance records often show only interval observation ranges; the other difficulty arises from the strong correlation structure, where component degradation links are clearly coupled with maintenance actions. The prediction model using independent feature assumptions exhibits unstable fluctuations during the parameter estimation stage, resulting in insufficient coverage of the prediction range, and maintenance scheduling relies on experience compensation.

[0127] When applying the method of this invention, the operation and maintenance data are first organized into a sample set, and time alignment and field standardization are performed on the sample set to generate structured records. The structured records include component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation intervals. Time alignment uses a unified time unit and a fixed sampling interval, mapping each record to an aligned timestamp. Field standardization ensures unified field naming and format. Component identifiers, fault mode identifiers, event identifiers, and maintenance action identifiers are uniformly encoded, keeping the encoded values ​​fixed within the sample set. The interval observation interval is represented by a lower bound and an upper bound, with the upper bound taking a finite value or an infinite upper bound. A knowledge graph is constructed based on the structured records, resulting in a node set, an edge set, and edge weights. The node set covers component nodes, fault mode nodes, event nodes, maintenance action nodes, operating condition nodes, and interval time nodes. The edge set covers the path from component nodes to fault mode nodes, fault mode nodes to event nodes, event nodes to maintenance action nodes, and so on. The association edges from component nodes to condition nodes and event nodes to interval time nodes are defined. The edge weights are determined by the edge type and the frequency of occurrence of the corresponding field in the structured record. Subsequently, a sampled parameter vector is established, and a mapping relationship between the sampled parameter vector and the node set is established. A graph correlation matrix is ​​generated based on the edge weights. A structured proposal distribution is constructed based on the graph correlation matrix. A parameter feasible region is established based on the interval time observation interval. Within the parameter feasible region, truncated proposal samples are generated to obtain candidate parameter vectors. The sample set is divided into subgraph data blocks according to the knowledge graph, and a subgraph log-likelihood cache is established. The candidate parameter vectors are used to calculate the log-likelihood increment. The delayed acceptance Metropolis–Hastings algorithm iteratively updates the sampled parameter vector under the constraints of the structured proposal distribution, subgraph log-likelihood cache, and log-likelihood increment. When the convergence condition is met, a posterior sample set is obtained. The posterior sample set is used to output the fault risk parameters and generate the fault interval time prediction results. The prediction results are expressed as the predicted mean, predicted median, lower bound of the prediction interval, and upper bound of the prediction interval.

[0128] Three baseline methods were set up for the comparative experiment. Baseline method A used the moving window averaging interval method, which statistically analyzed the window mean and window quantile of historical failure interval time samples according to component identification, and output the single-point predicted value and the empirical prediction interval. Baseline method B used a parametric survival model fitting, with the distribution form being Weibull distribution. The operating condition variable was included as a covariate in the scaling term, and the parameters were estimated using maximum likelihood estimation. No knowledge graph structure or delayed acceptance sampling was introduced. Baseline method C used a Cox proportional hazards model, with the operating condition variable included as a covariate in the hazard function. Partial likelihood estimation was used. No knowledge graph structure or interval consistency sampling was introduced. The dataset consisted of continuous collection cycles from 48 units, with 312,480 structured records, involving a total of 19,260 event identifiers and a total of 4,836 pairs of failure event records. The interval observation interval accounted for 37.9% of the interval observations, and the right censoring rate was 21.4%. The comparison indicators covered prediction accuracy, interval reliability, and computational cost. The evaluation objects selected were three types of high-frequency failure components: gearbox, generator, and pitch system. The comparison results are shown in Table 1:

[0129] Table 1. Quantitative Comparison of Different Methods in Failure Interval Prediction Task

[0130] The results show that baseline method A relies on approximate reduction for interval observation processing, underutilizes right-censored samples, and has low interval coverage. While baseline methods B and C introduce a survival analysis framework, they do not use knowledge graph relational structures in the parameter estimation process. Covariate correlations are represented as independent inputs in the model, leading to parameter estimation fluctuations under strong correlation conditions and making it difficult to achieve the nominal level of predicted interval coverage. The method in this invention constructs a structured proposal distribution through a knowledge graph correlation matrix and truncates proposal samples through the parameter feasible region, ensuring consistency between interval observations and right-censored observations during inference. The subgraph log-likelihood buffer and log-likelihood increment enable delayed acceptance of the Metropolis-Hastin method. The GS algorithm reduces redundant calculations during the update phase, and its single-round inference time is lower than that of baseline methods B and C. At the same time, it maintains a high interval coverage and a low negative logarithmic score. On the operation and maintenance side, maintenance plan windows are generated based on the lower and upper bounds of the predicted intervals. The window length is adaptively changed by subsequent a priori uncertainty. The proportion of maintenance records for gearbox and generator components completed within the planning window increases from 71.6% in baseline method B to 84.2% in the method of this invention, and the proportion for pitch system increases from 69.3% to 82.7%. The proportion of temporary maintenance records outside the planning window decreases from 27.8% to 15.1%. The number of cross-shift emergency repair work orders in the resource scheduling log decreases from an average of 32 per operating cycle to 19.

[0131] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the fault interval time of offshore wind turbines based on knowledge graphs, characterized in that, Includes the following steps: Step 1: Collect operation and maintenance data of offshore wind turbines and generate a sample set; Step 2: Perform time alignment and field standardization on the sample set to generate structured records. The structured records include component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation ranges. Step 3: Construct a knowledge graph based on structured records to obtain the node set, edge set, and edge weights; Step 4: Establish the parameter vector to be sampled, establish the mapping relationship between the parameter vector to be sampled and the node set, generate the graph correlation matrix based on the edge weights, and construct the structured proposal distribution based on the graph correlation matrix; Step 5: Establish the feasible region of parameters based on the observation interval, generate truncated proposal samples within the feasible region of parameters, and obtain the candidate parameter vector; Step 6: Divide the sample set according to the knowledge graph to obtain subgraph data blocks, establish a subgraph log-likelihood cache, and calculate the log-likelihood increment based on the candidate parameter vector; Step 7: Using the delayed acceptance Metropolis–Hastings algorithm, based on the structured proposal distribution, subgraph log-likelihood buffer, and log-likelihood incremental iterative update of the parameter vector to be sampled, the posterior sample set is obtained when the convergence condition is met. Step 8: Output fault risk parameters based on the posterior sample set and generate fault interval time prediction results.

2. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step two specifically involves: Time alignment is performed on the sample set. The time alignment uses a uniform time unit and a preset sampling interval to map each record in the sample set to an aligned timestamp. Perform field standardization on the sample set. Field standardization includes unifying field naming and field format. Extract component identifiers, fault mode identifiers, event identifiers, operating condition variables, maintenance action identifiers, and interval observation ranges from the sample set to form a structured record; Component identifiers, fault mode identifiers, event identifiers, and maintenance action identifiers are uniformly coded, and a one-to-one correspondence is established between the coded values ​​and the field values, while maintaining consistency in the sample set; The unit of operation condition variables is standardized and missing, and the values ​​of the operation condition variables are generated and written into structured records. The structured records are sorted based on the aligned timestamps. Fault event record pairs are determined according to the component identifier and event identifier. The time difference between the fault event record pairs is calculated, and the lower and upper bounds of the interval observation interval are determined and written into the structured records.

3. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step three specifically involves: Based on the structured records, component identifiers, fault mode identifiers, event identifiers, and maintenance action identifiers are extracted. Based on the extraction results, component nodes, fault mode nodes, event nodes, and maintenance action nodes are generated and aggregated to form a node set. The edge set is generated based on the structured record. The edge set includes the associated edges from component nodes to fault mode nodes, the associated edges from fault mode nodes to event nodes, and the associated edges from event nodes to maintenance action nodes. Each associated edge records the starting node identifier, the ending node identifier, and the edge type. Operating condition nodes are generated based on structured records. The operating condition nodes are determined by the field names of operating condition variables and the values ​​of the operating condition variables are recorded. The associated edges from event nodes to operating condition nodes are generated and written into the edge set. Based on structured records, interval time nodes are generated. Each interval time node records the lower and upper bounds of the observation interval. The associated edges from event nodes to interval time nodes are generated and written into the edge set. Edge weights are calculated based on the edge set. The edge weights are determined according to the edge type and the frequency of occurrence of the corresponding field in the structured record. The edge weights are written into the edge set. A knowledge graph is constructed based on the node set, edge set, and edge weight, and the node set, edge set, and edge weight are output.

4. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step four specifically involves: A parameter vector to be sampled is established. The parameter vector is formed by arranging the parameter components in a fixed order. The number and order of the parameter components remain unchanged during the iteration process. Establish a mapping relationship between the vector of parameters to be sampled and the set of nodes, where each parameter component is assigned a node identifier in the set of nodes. A node relevance table is generated based on the edge set and edge weights. The node relevance table uses each pair of node identifiers in the node set as an index. The node relevance is the sum of the edge weights of the matching node identifiers of the starting node identifier and the ending node identifier in the edge set. The node relevance is zero when the edge set does not contain any matching node identifier pairs. Based on the mapping relationship, the node correlation table is projected onto the parameter component pairs to generate the graph correlation matrix. The row index and column index of the graph correlation matrix correspond to the order of the parameter components of the parameter vector to be sampled. The elements of the graph correlation matrix take the node correlation of the parameter component pairs mapped to the node identifier pairs. A structured proposal distribution is constructed based on the graph correlation matrix. The structured proposal distribution adopts a multivariate normal distribution. The mean vector of the structured proposal distribution is taken as the current value of the parameter vector to be sampled. The covariance matrix of the structured proposal distribution is determined by the graph correlation matrix and the step size coefficient.

5. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step five specifically involves: The time interval observation interval is read from the structured record. The time interval observation interval is represented by a lower bound value and an upper bound value. The upper bound value can be a finite value or an infinite upper bound. Establish a calculation rule for the interval time distribution corresponding to the parameter vector to be sampled. The interval time distribution calculation rule is used to calculate the interval probability that the interval time falls into the interval time observation interval under the given parameter vector to be sampled. The interval probability is calculated one by one based on the structured records. The interval probability is the probability that the interval time is not less than the lower bound and not greater than the upper bound. Under the condition that the upper bound is infinite, the interval probability is the probability that the interval time is not less than the lower bound. The feasible region of parameters is established based on the interval probability. The feasible region of parameters includes the values ​​of the sampled parameter vector where all interval probabilities are positive. Proposal samples are generated based on the structured proposal distribution. Proposal samples that fall within the parameter feasible region are identified as truncated proposal samples, and these truncated proposal samples are identified as candidate parameter vectors.

6. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step six specifically involves: Based on the connectivity of node sets in the knowledge graph, a subgraph identifier is determined for each structured record. The subgraph identifier is determined by the node identifiers of the component identifier, fault mode identifier, event identifier, and maintenance action identifier contained in the structured record in the node set. Structured records are grouped according to subgraph identifiers to form subgraph data blocks. Each subgraph data block contains a subgraph identifier and a set of structured records. Establish a subgraph log-likelihood cache. The subgraph log-likelihood cache records the log-likelihood value of the subgraph data block under the current value of the parameter vector to be sampled, using the subgraph identifier as the index. The set of parameter change components is determined based on the candidate parameter vector and the parameter vector to be sampled. The set of parameter change components is determined by the parameter components whose values ​​are different from those of the parameter components in the candidate parameter vector and the parameter components in the parameter vector to be sampled. Based on the mapping relationship between the parameter vector to be sampled and the node set, the parameter change component set is mapped to the change node identifier set, and the affected subgraph identifier set is determined based on the change node identifier set; For each subgraph identifier in the affected subgraph identifier set, calculate the log-likelihood value of the subgraph data block under the conditions of the candidate parameter vector values, read the log-likelihood value corresponding to the subgraph identifier in the subgraph log-likelihood cache, and use the difference to calculate the log-likelihood increment.

7. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step seven specifically involves: Read the current value of the parameter vector to be sampled, the structured proposal distribution, and the parameter feasible region. Generate proposal samples based on the structured proposal distribution. When a proposal sample satisfies the parameter feasible region condition, it is determined as a candidate parameter vector. The proposal density ratio is calculated based on the structured proposal distribution. The proposal density ratio is obtained by dividing the proposal density of the candidate parameter vector under the condition of the parameter vector to be sampled by the proposal density of the parameter vector to be sampled under the condition of the candidate parameter vector. Calculate the subgraph weight corresponding to the subgraph identifier in the affected subgraph identifier set based on the edge set and edge weight, set a subgraph weight threshold, and obtain a pre-filtered subgraph identifier set from the affected subgraph identifier set based on the subgraph weight threshold; The Metropolis–Hastings algorithm with delayed acceptance is used to perform pre-screening acceptance decision. The pre-screening acceptance decision is based on the ratio of the correspondence likelihood increment to the proposal density of the pre-screened subgraph identifier set. When the pre-screening acceptance criteria are met, the precise acceptance criteria are executed. The precise acceptance criteria generate the precise acceptance probability based on the ratio of the log-likelihood increment to the proposal density for the affected subgraph identifier set. When the precise acceptance criteria are met, the candidate parameter vector is written into the parameter vector to be sampled and the log-likelihood cache of the subgraph corresponding to the affected subgraph identifier set is updated. Repeatedly execute candidate parameter vector generation, pre-screening acceptance determination, and precise acceptance determination to obtain the posterior sample set when the convergence condition is met.

8. The method for predicting the fault interval time of offshore wind turbines based on knowledge graphs according to claim 1, characterized in that, Step eight specifically involves: Read the posterior sample set, which contains the sample values ​​of the parameter vector to be sampled; Fault risk parameters are calculated based on the posterior sample set. The fault risk parameters include the statistical values ​​of the parameter components of the sampled parameter vector, and the statistical values ​​of the parameter components include the mean, variance, and quantile. Read the interval time distribution calculation rule, input the sample values ​​of the parameter vector to be sampled in the posterior sample set into the interval time distribution calculation rule, and generate the fault interval time prediction distribution; The fault interval time prediction results are calculated based on the fault interval time prediction distribution. The fault interval time prediction results include the prediction mean, prediction median, lower bound of the prediction interval, and upper bound of the prediction interval. Output the failure risk parameters and failure interval prediction results.

Images