An explainable industrial fault diagnosis method based on knowledge graph multi-hop reasoning

Abstract

The application provides an explainable industrial fault diagnosis method based on knowledge graph multi-hop reasoning, mainly aiming at a wind turbine fault diagnosis scene, and is used for solving problems of candidate path noise, insufficient evidence readability and low answer credibility. The method is based on a wind turbine fault diagnosis knowledge graph, performs anchor entity recognition and constrained multi-hop path retrieval on an input question, obtains a candidate reasoning path set, rearranges and screens high-quality paths through semantic feature and structural feature fusion, and generates natural language evidence. The method further aggregates candidate answers, verifies consistency and corrects backtracking, and finally outputs a diagnosis answer, a supporting path and evidence description. The application can improve the accuracy, explainability and credibility of industrial fault diagnosis, and provides technical support for intelligent operation and maintenance of complex industrial equipment.

Description

Technical Field

[0001] This invention relates to the fields of intelligent operation and maintenance, knowledge graph question answering, and fault diagnosis technology for wind power generation. Specifically, it relates to an interpretable industrial fault diagnosis method based on knowledge graph multi-hop reasoning, which is mainly aimed at the fault diagnosis scenario of wind turbine generator sets. It can be used for fault location, root cause analysis, traceability of detection basis, and auxiliary decision-making for handling of five key components: rotor, main shaft, bearing, gearbox, and generator. Background Technology

[0002] Wind turbine generators operate under conditions of variable speed, variable load, and complex environments for extended periods. Structural coupling and fault propagation relationships exist among the five key components: rotor, main shaft, bearings, gearbox, and generator. An anomaly in a single component can propagate along mechanical transmission or functional linkages; therefore, diagnostic methods capable of pinpointing the cause of a fault and providing traceable evidence are needed.

[0003] Fault diagnosis of wind turbine generators involves seven categories of heterogeneous knowledge: equipment structure, component status, fault phenomena, fault causes, detection methods, remedial measures, and fault impacts. This knowledge is scattered across equipment manuals, maintenance records, fault cases, and operation and maintenance experience. The same fault phenomenon may also correspond to multiple cause chains or remedial chains, making the diagnostic process require multi-hop reasoning by combining multiple entities and relationships.

[0004] Knowledge graphs can organize knowledge about wind turbine generator failures using triples of entities, relationships, and facts. They explicitly represent the relationships between equipment or components and failure phenomena, causes, detection methods, handling measures, and impacts, providing a structured foundation for failure propagation analysis, multi-hop path reasoning, and diagnostic basis tracing.

[0005] However, the knowledge graph of wind turbine generator faults contains a variety of relationship types. When bearings or gearboxes become highly connected nodes, multi-hop expansion can easily generate a large number of candidate paths. If expansion is performed directly without constraints, paths that appear related but lack fault mechanism support will accumulate rapidly, increasing computational load and interfering with the correct fault chain ordering.

[0006] Candidate paths typically exist in the form of structured triple sequences, which are easy for machines to process but not easy for maintenance personnel to directly review. Simply outputting candidate answers or confidence levels cannot fully explain how the fault phenomenon is deduced from the component and cause relationships to the diagnostic conclusion, and it is also difficult to meet the requirements of interpretability, verifiability, and traceability.

[0007] Furthermore, the path with the highest ranking score is not necessarily related to the problem. Figure 1 Root cause queries may yield high-scoring candidates for handling measures or failure impacts. However, without posterior verification of the semantic consistency of relation types, path structures, and "question-answer-evidence" relationships, answer type bias is likely to occur.

[0008] In summary, existing technologies face at least three critical issues that urgently need to be addressed:

[0009] • The candidate paths are subject to strong noise interference. The knowledge graph of wind turbine generator faults contains many component associations and causal chains. Simple neighborhood expansion can easily introduce paths that are inconsistent with the current problem's intent, affecting subsequent path rearrangement and answer aggregation.

[0010] • The readability and traceability of reasoning evidence are insufficient. Structured paths are difficult to directly translate into diagnostic criteria that operations and maintenance personnel can understand, and simply outputting answers or scores is insufficient to explain the derivation process of diagnostic conclusions.

[0011] • The answer credibility verification mechanism is imperfect. High-scoring candidates may belong to the causes of failure, detection methods, handling measures, or impacts of failure. If there is a lack of consistency verification between the problem intent, path relationship, and supporting evidence, the initial answer is difficult to reliably correct.

[0012] To address the aforementioned issues, this invention proposes an interpretable industrial fault diagnosis method based on knowledge graph multi-hop reasoning. This method integrates candidate path retrieval, semantic structure fusion and rearrangement, path evidence generation, answer aggregation, consistency verification, and backtracking correction into a closed-loop process, ensuring that the diagnostic results simultaneously include the answer, supporting paths, and readable evidence. Summary of the Invention

[0013] This invention proposes an interpretable industrial fault diagnosis method based on knowledge graph multi-hop reasoning. First, anchor entities corresponding to wind turbine generator components or fault phenomena are identified from the fault problem, and relational constraints are formed based on the diagnostic intent of cause, detection, handling, or impact. Then, constrained multi-hop path retrieval is performed starting from the anchor entities. Subsequently, the structure features of the problem semantics and fault propagation paths are combined to rearrange the high-quality structured paths into natural language evidence. Next, candidate answers are formed based on the path endpoints, and their supporting paths and evidence are aggregated. Finally, structured and semantic consistency verification is performed; when the answer type or evidence chain is inconsistent, a reliable result is selected from the candidate answers through backtracking. Detailed steps are shown below:

[0014] (1) Basic Representation and Candidate Path Retrieval: First, a knowledge graph representation of wind turbine generator fault diagnosis is established, and candidate paths are retrieved based on this representation. In the retrieval stage, search depth, node expansion scale, and single relation type expansion constraints are introduced to control the number of candidate paths and reduce noise caused by high-connection components and high-frequency relations.

[0015] 1) Knowledge Representation and Path Definition. The knowledge graph for wind turbine generator fault diagnosis is denoted as a directed heterogeneous graph. .picture The entity types include wind turbine generator equipment entities; component entities include rotors, main shafts, bearings, gearboxes, and generators; diagnostic entities include fault phenomena, fault causes, detection methods, remedial measures, and fault effects. Relationship types include composition, association, cause, detection, remediation, and impact. Composition and association relationships are used for path connections between equipment, components, and fault phenomena. Cause, detection, remediation, and impact relationships correspond to fault cause, detection method, remedial measure, and fault impact diagnostic results, respectively. Facts in the graph are recorded in a triplet format of "head entity—relationship—tail entity".

[0016] The natural language wind turbine generator faults to be diagnosed are denoted as follows: The system first identifies one or more anchor entities, and then uses these anchor entities as the starting point of the path, expanding hop-by-hop along the directed relation edges in the knowledge graph. Any multi-hop reasoning path is represented as follows:

[0017] (1)

[0018] in, This represents a path composed of entities and relations arranged in a logical order. This represents the number of relation edges contained in the path, i.e., the path hop count. For any Its value ranges from 1 to , No. Jump from the head entity ,relation Tail-end entity Composition, and the jump corresponds to the diagram. A fact triple in the middle; Represents the anchor point entity at the start of the path. Represents the entity at the end of the path.

[0019] Subsequent path retrieval, structural feature calculation, answer merging, and evidence output are all based on the above path representation.

[0020] 2) Constrained Path Expansion. After completing the basic definition, the system performs a constrained breadth-first search starting from the set of anchor entities. The purpose of this process is to prioritize recalling paths that may support the answer, while avoiding path explosion caused by high-frequency relationships or highly connected nodes in complex graphs. The initial candidate path set is represented as follows:

[0021] (2)

[0022] in, This represents the knowledge graph for fault diagnosis of the same wind turbine generator set mentioned above; This indicates a constrained breadth-first search; Indicates from the question The set of anchor point entities identified in the graph, whose elements are graphs. Entity nodes in; Indicates the maximum search depth; This indicates the maximum number of candidate edges that can be retained for each node; This represents the maximum number of candidate edges that can be retained for each relation type under the same node; This represents the initial set of candidate paths obtained under the above constraints.

[0023] Each expansion consists of two steps. The first step is to sort the candidate edges according to their priority in equation (3) from highest to lowest, and retain at most [number missing] edges. The first step is to select candidate edges. The second step is to group them by relation type, ensuring that the number of edges retained for each relation type does not exceed a certain limit. The search depth reached Alternatively, stop expanding when the current layer no longer generates new paths.

[0024] , and All are pre-defined positive integers and satisfy the following conditions: Not greater than The constraints imposed by the three on the search process are... Pass it on to the next step.

[0025] Problem constraints are identified based on relation names, relation aliases, and diagnostic question prompts: cause or root cause corresponds to cause-type relations and fault cause-type endpoints; detection or inspection corresponds to detection-type relations and detection method-type endpoints; handling, repair, or treatment corresponds to handling-type relations and handling measure-type endpoints; impact or consequence corresponds to impact-type relations and fault impact-type endpoints; no relation signature constraints are applied when no relation prompt is identified; and only the consistency between the path endpoint and the candidate answer is checked when no endpoint role is identified.

[0026] The maximum search depth is determined by combining the typical levels of the wind turbine generator fault propagation chain and the recall of candidate paths. The single node expansion limit is determined by combining the out-degree of the graph node and the computational cost. The single relation type expansion limit is used to suppress the excessive expansion of high-frequency relations. The specific configuration is shown in Table 1.

[0027] 3) Edge Priority Evaluation and Path Selection. During each hop expansion, the system does not retain all adjacent edges. Instead, it evaluates candidate edges based on question terms, target entities, relational hints, and semantic similarity, prioritizing relations more aligned with the question's intent. Candidate edge priority is expressed as follows:

[0028] (3)

[0029] in, This indicates the hop number of the current path, the hop number. Jump candidate edges are ; Indicate the problem The priority score of the candidate edge for a single hop under the given conditions only evaluates the current hop and does not represent the score of the entire path. This represents the set of all terms after the problem is normalized. This represents a set of terms indicating the relationship between the question and the answer. This indicates text normalization and term segmentation. Indicates the number of elements in the set. This indicates that the entity name will be returned. To represent the intersection of sets, This indicates string concatenation. The sentence vector cosine similarity between the question text and the concatenated text of the relational text and the target entity name; , , and The non-negative weights control relation term matching, target entity term matching, relation hint matching, and semantic similarity, respectively. This score is used for hop-by-hop candidate edge selection in Equation (2) and as a hop-by-hop summation term in Equation (4).

[0030] After completing the multi-hop expansion, the system obtains the overall priority of the path based on the priority of each edge within the path, and retains the candidate paths with the highest global ranking:

[0031] (4)

[0032] (5)

[0033] in, Indicate candidate path The overall retrieval priority is calculated from the score of each jump candidate edge by equation (3). to The scores are obtained by summing them up sequentially. When the path contains only one hop, the score of the path is equal to the edge score of that hop obtained according to formula (3); when the path contains two or more hops, the score of the path is the sum of the edge scores of each hop. This means retaining a specified number of elements based on their given scores, from highest to lowest. Indicates the number of globally reserved paths. Indicates from The most retained The set of candidate search paths obtained after searching for paths; the number of candidate paths is less than All paths are preserved at that time.

[0034] (2) Path rearrangement and evidence generation

[0035] Candidate path retrieval focuses on recall and may still include paths that appear relevant but lack sufficient causal logic for the failure. To further filter high-quality evidence paths, the re-ranking stage utilizes both semantic matching features and structural statistical features, and transforms the retained structured paths into readable evidence.

[0036] 1) Semantic and structural feature fusion. For the retrieval candidate path set obtained by equation (5) The first in Candidate paths The system will display the problem text. and path text Inputting the same sentence into the vector encoder yields the semantic representations of both sentences:

[0037] (6)

[0038] in, Indicate the problem semantic vector; Indicates the first Candidate paths semantic vectors, This represents the path text composed of entity names and relation names in the path, arranged in the path order. This represents a sentence vector encoder with frozen parameters. Both types of vectors are generated by the same sentence vector encoder and therefore have the same dimension.

[0039] (7)

[0040] in, Indicates the first The semantic interaction feature vectors of the candidate paths; each vector is concatenated along the feature dimension in the order listed in equation (7); This means taking the absolute value of the difference between corresponding elements of two vectors; This represents element-wise multiplication. After concatenating the four parts, the dimension of this vector is four times that of a single sentence vector.

[0041] (8)

[0042] in, Indicates the candidate path number. Indicates the first Candidate paths The nine-dimensional structural eigenvectors; Indicates the number of hops in the path; This indicates the initial search ranking, starting from 1. Indicates the reciprocal of that ranking; Indicates the number of terms in the path text; The number of common terms between the question and the complete path text; The number of common terms representing the entity names of the problem and the path endpoint, where The endpoint entity of the path; Indicates the number of different relation types in the path; Indicates whether a pattern node exists in each hop target node of the path; The value is 1 if the path endpoint is a pattern node and 0 otherwise. The nine structural features in Equation (8) form a structural feature vector in the order listed and serve as the input to the structural branch of Equation (9).

[0043] All nine components are calculated directly from the current question and candidate paths: the path hop count is the number of relation edges, the initial search ranking starts counting from 1, the reciprocal of the ranking is obtained from that ranking, the path text length and the overlap of the two items are counted according to normalized terms, and the relation diversity is counted according to different relation types; the path-level pattern node indicator is 1 when any target node matches the pattern identifier pre-registered in the knowledge graph, otherwise it is 0, and the endpoint pattern node indicator only checks the path endpoint.

[0044] Semantic and structural features are respectively processed in their respective branches for representation transformation, and then fused at the output layer to obtain the path relevance score.

[0045] (9)

[0046] in, and These represent the feature transformation functions for semantic branches and structural branches, respectively. and These represent the output vectors of the two branches, and the two output vectors are concatenated in the listed order. and These represent the weight vector and bias of the output layer, respectively. Take the transpose during calculation. Indicates the first The reordering score of each path. The trainable parameters mentioned above are determined by minimizing the pairwise ranking loss of equation (22), while the sentence vector encoder parameters remain frozen. This reordering score has a different meaning from the path priority in the retrieval stage of equation (4). Descending order retains at most The path is given, and the set of rearranged paths is obtained. And use it for candidate answer formation and path evidence aggregation in Equation (13).

[0047] 2) Path Linearization and Evidence Generation. The high-quality paths retained after rearrangement are still structured triple sequences, which are not convenient for manual review. Therefore, the system first decomposes the paths into ordered triples and adds structure tags to obtain linearized path text:

[0048] (10)

[0049] in, Indicates the candidate path number; Indicates the first The hop number within the path ranges from 1 to... ; , and The first one is formed in sequence. The jump's head entity, relation, and tail entity; Representing a path Structured linear text; "triples" The arrows and symbols are used to distinguish the jumps and their connection order. according to Concatenate the triplets in ascending order.

[0050] Candidate paths are formed by hopping from anchor entities, combined with relational constraints identified in the current fault problem. The relational types and number of hops in the path are determined according to the specific fault problem. For any retained candidate path, the system linearizes each jump-head entity, relation, and tail entity sequentially according to the actual jump order; each jump relation is determined based on the actual directed connections of composition, association, cause, detection, handling, and impact relationships in the knowledge graph.

[0051] Subsequently, the evidence generation model transcribes the linearized path into a natural language evidence description:

[0052] (11)

[0053] in, This indicates a sequence-to-sequence evidence generation model using the T5-small architecture; This represents the set of trainable parameters for the model, including encoder, decoder, and output layer parameters, and is determined by minimizing the generated negative log-likelihood loss of equation (23). This represents the structured linear path text obtained according to equation (10); Indicates by the first Natural language evidence text generated from the path.

[0054] (12)

[0055] in, Indicates the first Comprehensive evidence corresponding to each path; This is an indicator of the validity of evidence, and its value is either 0 or 1. Represents structured linear path text; This represents natural language evidence text. The generated text is non-empty, contains the names of the first and last entities in the path, and has a character count not exceeding [a certain value]. hour, Set the value to 1, and retain both types of text; otherwise, set the value to 0, and revert to structured linear path text.

[0056] The maximum number of characters for evidence validity is determined by considering the completeness and redundancy of evidence in the verification set, and is specifically configured as 300 characters. This character limit is used for evidence verification after generation, and is used to constrain different stages, respectively, along with the maximum number of newly added tokens in the decoding stage.

[0057] (3) Answer aggregation, consistency verification and backtracking correction

[0058] Different candidate paths may lead to the same answer, or they may lead to maintenance measures, failure consequences, or reasons for overly broad granularity. The system first aggregates the path support pointing to the same answer, and then combines the results of structured consistency and semantic verification to determine whether the current answer is reliable; if it is inconsistent with the question intent or the chain of evidence, a backtracking correction is performed.

[0059] 1) Answer aggregation. The system will rearrange and retain the set of candidate paths. By deduplicating the endpoint entities of each path, a set of candidate answers is obtained. Then, the supporting paths and evidence corresponding to each candidate answer are merged according to the path endpoint entity:

[0060] (13)

[0061] in, This represents the set of candidate paths retained after rearrangement. Indicates the first Path, Representing a path The final entity, Indicates the first 10 candidate answers Indicates the endpoint entity is The set of supporting paths; Representing a path The comprehensive evidence formed according to formula (12), Indicates and The path in the equation corresponds to a comprehensive set of evidence. Equation (13) merges supporting paths and evidence according to the path endpoints. The conditions after the vertical lines are used to limit the members of the set, and the double vertical lines indicate the number of elements in the set. This equation is a set construction relationship and does not represent probability. Candidate answer set Depend on The endpoint entity is obtained by deduplication, so each candidate answer corresponds to at least one supporting path.

[0062] (14)

[0063] in, Indicates candidate answers The aggregate score; This represents the positive class matching probability output by the answer-level rearranger according to equation (25); , and These represent the number of supporting paths, the highest supporting path reorder score, and the average supporting path reorder score, respectively, within the candidate answer set for the same question after min-max normalization. If the maximum and minimum values ​​of a certain quantity are the same among all candidate answers, then its normalization result is uniformly set to 1. This indicates the ranking of the best support path, starting from 1. The reciprocal item indicates the reciprocal of this ranking, and its value is greater than 0 and does not exceed 1. to The weights are non-negative and sum to 1, with each of the five weights set to 0.2. The aggregated score is used in equation (15) to select the initial answer and forms the comprehensive verification ranking score in equation (20).

[0064] (15)

[0065] Equation (15) is derived from the set of candidate answers. Selected from The largest candidate answer; This serves as the initial diagnostic answer obtained from the above, and as the current answer for semantic verification in equation (19) and final answer selection in equation (21). When aggregate scores are the same, the answer with priority is selected. The smaller candidate answer; if they are still the same, the original candidate ranking is maintained.

[0066] 2) Structured Consistency Verification. To avoid the answer being determined solely by local high-scoring paths, the system performs posterior checks on three aspects: evidence support strength, relation signature consistency, and path structure consistency. The three types of structured consistency signals are calculated as follows:

[0067] (16)

[0068] (17)

[0069] (18)

[0070] in, Indicates candidate answers Strength of evidence; molecular accumulation The scores for each non-negative reordering path are calculated by summing the denominators. Scores for all non-negative multiple permutations of all paths; Used to truncate negative scores to 0. To prevent positive smoothing constants with denominators of 0. Wherein, This represents the set of relation signatures obtained from the problem relation hint mapping. Indicates candidate answers The set of relation types contained in the supporting path. Representing a path No. The relationship of jumping, This indicates the number of hops in the path; Indicates consistency of relation signatures, when When not empty, take the number of elements in the intersection of the two sets and The ratio, when When the value is 1, then... , and These are used to check the path direction and continuity, the number of hops in the path, and the path's endpoint and its role. A value of 1 is returned if the condition is met, and 0 otherwise. This represents the structural matching value of a single support path; This represents the average value of all supporting path structure matching values. All three types of consistency results are used as inputs to equation (20).

[0071] Direction and continuity indicators are based on the path starting from the anchor point entity, the continuity of two adjacent jump triples, and the relationship direction with the graph. The hop count indicator is set to 1 when the hop count falls within the expected range of the problem; the endpoint indicator is set to 1 when the endpoint equals the candidate answer and, if the diagnostic intent has been identified, belongs to the entity type corresponding to that intent (i.e., fault cause, detection method, remedial action, or fault impact); otherwise, it is set to 0. When the hop count is not explicitly stated in the problem, the expected range is 1 to 3; when the diagnostic intent has not been identified, only the consistency between the endpoint and the candidate answer is checked. The evidence support smoothing constant is set to 1 × 10⁻. 8 This is only used to avoid a denominator of 0.

[0072] 3) Semantic Validation and Backtracking Correction. The structured consistency signal provides computable constraints, and the semantic validation model further determines whether there is consistency between the question, the current answer, and the evidence text. The system parses the raw output of the semantic validation model into structured validation results:

[0073] Before semantic verification, the system sorts the candidate answers in descending order of aggregate score and calculates the score difference between the top two candidate answers. When the score difference is less than the two-candidate arbitration threshold, the system fine-tunes the large language model by freezing the input parameters of the top two candidate answers, their supporting paths, comprehensive evidence, and three types of structured consistency values. Otherwise, only the first one is input. The input template contains the question text, candidate answer, supporting path, evidence support strength, relation signature consistency, path structure consistency, and comprehensive evidence in sequence, and requires the model to output only the two fields "best_answer" and "reason". The system parses the original output into the "support", "correction", or "undetermined" state as shown in Equation (19) based on the completeness of the fields, whether the answer belongs to the input candidate set, and whether the corresponding supporting path is not empty. When the fields are missing, the output answer does not belong to the candidate set, the corresponding supporting path is empty, or the parsing fails, it is recorded as "undetermined", and backtracking selection is performed by Equation (20) and Equation (21).

[0074] The two-candidate arbitration threshold is used to compare the top two candidates: if the score difference is less than the threshold, both candidates are verified at the same time; otherwise, only the first candidate is verified. The threshold is determined by combining the answer selection effect on the verification set, and the specific configuration is 0.005.

[0075] (19)

[0076] in, This indicates that the system outputs the structured verification results generated by parsing based on the semantic verification model; Choose "Support", "Revise", or "Undetermined"; This represents the candidate answer selected by the semantic verification model, and can be empty in the "undecided" state; Natural language descriptions used to express judgments.

[0077] (20)

[0078] in, Represents the set of candidate answers to the same question. Internally, for aggregate scores The result obtained after performing exponential normalization; the hat symbol indicates the normalized quantity. In the summation, the first... 10 candidate answers; Indicates candidate answers The comprehensive verification ranking score, , , and represents the fusion weights for normalized aggregation score, evidence support strength, relation signature consistency, and path structure consistency, respectively; all are non-negative and sum to 1. In the actual calculation of exponential normalization, the maximum aggregation score for the current problem is subtracted from each exponential term to avoid numerical overflow; this processing does not change the normalization result. Equation (21) is used to perform backtracking selection when semantic verification does not directly determine the answer.

[0079] (twenty one)

[0080] in, This represents the corrected answer proposed by semantic verification. This indicates the status obtained after performing a second verification on the corrected answer. When the semantic verification status is "Support," Get the initial answer When the semantic verification status is "corrected", the system first checks... and A second confirmation will only be performed after the above checks have passed; if the checks fail or the confirmation fails, an order will be issued. "Undetermined". For "correction" and When "support", Pick In other cases, from Select the comprehensive verification ranking score The largest candidate answer. Confirm. Then, output the answer along with its corresponding set of supporting paths and a comprehensive set of evidence.

[0081] This mechanism enables the system to backtrack from the set of candidate answers and select a more reliable diagnostic result when the initial answer does not conform to the question's intent or the chain of evidence is insufficient.

[0082] (4) Model training and online inference strategies

[0083] This invention employs a phased training strategy: the path rearrangement model, the evidence generation model, and the answer-level rearranger are trained separately, while the sentence vector encoder and the semantic verification large language model remain frozen during the online reasoning phase.

[0084] 1) Path Reordering Model Training. The path reordering model is used to learn the relative ordering relationship that "the correct path should precede the noisy path". For each question in the training set, the system labels positive and negative paths based on the standard answer entity and optimizes the path scores using pairwise ordering loss:

[0085] (twenty two)

[0086] in, Represents the positive path under the same problem With negative path Pair sorting loss; and These represent the rearrangement scores for positive and negative paths, respectively. This indicates the pre-defined sorting interval; specific values ​​are shown in Table 1. This indicates taking the larger value between 0 and the value within the parentheses.

[0087] Positive paths are those where the endpoint entity belongs to the standard answer set; negative paths are those where the endpoint entity does not belong to the standard answer set. Negative paths are divided into high-ranking difficult negative paths and other medium-difficulty negative paths according to their search ranking, and are paired with positive paths to form training sample pairs; the number of positive and negative paths extracted for each question is shown in Table 1. Formula (22) is calculated for all positive and negative paths of the same question and the average is taken. The average loss is used for one parameter update. The AdamW optimizer is used to train the path reordering model. The learning rate and number of training rounds are shown in Table 1, and the model parameters determined at the end of training are saved.

[0088] 2) Training the evidence generation model. The evidence generation model is used to learn the mapping from structured paths to natural language evidence.

[0089] Training samples are constructed from problems with standard path annotations. Each standard path is converted into structured linear text according to Equation (10) and used as model input; each jump-head entity, normalized relation text, and tail entity are sequentially combined to form a statement, which serves as target evidence. Repeated input-target pairs are retained only once.

[0090] (twenty three)

[0091] in, Represents the target evidence word sequence; Indicates the total number of target lexical units; Indicates the current position; Indicates the first One target word element; Indicates the first to the second Given a known target word, when This indicates an initial condition that includes only the sequence start marker; The parameter is The conditional probabilities given by the evidence generation model; This represents the sum of the negative log-likelihoods at each position. The evidence generation model uses T5-small, and the learning rate, number of training rounds, batch size, beam search width, and input / output length are shown in Table 1. The threshold for the number of characters for evidence validity specified in Equation (12) and the upper limit of the word count in the decoding stage are two independent constraints; even if the generated result does not reach the upper limit of the word count, as long as the number of characters exceeds the threshold, it will still revert to structured path text.

[0092] 3) Answer-level rearranger training. The answer-level rearranger is used to determine whether candidate answers are consistent with the standard answer. For candidate answers... Construct a ten-dimensional answer-level feature vector The ten components are, in order: the number of overlapping terms between the question and the candidate answer, the overlap ratio, the number of terms in the question, the number of terms in the candidate answer, the number of supporting paths, the highest reorder score of the supporting path, the average reorder score of the supporting path, the best supporting path ranking, the last supporting path ranking, and whether there is a non-empty supporting path.

[0093] (twenty four)

[0094] in, Indicates candidate answers The corresponding ten-dimensional answer-level feature vector; the ten components represent, in turn, the number of normalized terms overlapping between the question and the candidate answer, the overlap ratio, the number of terms in the question, the number of terms in the candidate answer, the number of supporting paths, the highest re-ranking score of the supporting path, the average re-ranking score of the supporting path, the best supporting path ranking, the last supporting path ranking, and the supporting path existence identifier. exist The value is 1 if the condition is met, and 0 otherwise; the ten components form a column vector in the order listed.

[0095] (25)

[0096] in, and These represent the weight vector and bias of the answer-level logistic regression model, respectively. Represents a logical function; A binary variable representing the correctness of candidate answers. This indicates that the candidate answer is consistent with the standard answer. This indicates inconsistency; Indicates that given a feature vector Under the given conditions, the predicted probability that a candidate answer is the correct answer. The predicted probability is denoted as .

[0097] The answer-level rearranger is trained using the binary cross-entropy loss shown in the following formula:

[0098] (26)

[0099] in, Indicates the first Ten-dimensional answer-level feature vectors of training samples. This represents the predicted probability that the sample belongs to the positive class. Represents a true binary label; This represents the training objective of the answer-level rearranger. This represents the total number of training samples. It is determined by minimizing equation (26). and After training, the parameters are fixed, and the positive class probability of the candidate answer is calculated according to Equation (25) during the online inference stage. In the implementation, a class-balanced logistic regression classifier is used, with a maximum number of iterations of 1000 and a random seed of 42.

[0100] After offline training, the weight vector and bias of the answer-level logistic regression model are saved. During online inference, the parameters are loaded, and the correct answer prediction probability shown in Equation (25) is calculated based on the ten-dimensional feature vector of each candidate answer. This probability is then used as the first term in Equation (14) for answer aggregation. The answer-level rearranger parameters are not updated during the online inference stage.

[0101] 4) Online Inference Strategy. During the online inference phase, each training module loads fixed parameters and no longer performs gradient updates. The system sequentially completes anchor entity recognition, candidate path retrieval, path rearrangement, evidence generation, answer aggregation, structured consistency verification, semantic verification, and backtracking correction. Semantic verification directly outputs the current answer; when semantic verification provides a corrected answer and the second confirmation status is "supported", the corrected answer is adopted; in other cases, the candidate answer is backtracked and selected according to the verification ranking score in equation (20). Finally, the selected answer and its supporting path set and evidence set are output.

[0102] The main beneficial effects of this invention include:

[0103] (1) This invention proposes a knowledge graph multi-hop reasoning and interpretable diagnosis process for wind turbine generator fault diagnosis, which unifies candidate path retrieval, semantic structure fusion and rearrangement, path to text evidence generation, answer aggregation and backtracking correction in the same framework.

[0104] (2) This invention transforms structured paths into natural language evidence, so that the diagnostic results not only contain answers, but also readable and traceable evidence chains.

[0105] (3) The present invention introduces a structured consistency verification and semantic verification mechanism to perform posterior checks on the answer, question and supporting evidence, and performs backtracking correction when the initial answer does not meet the question intent.

[0106] (4) In the training phase, the present invention optimizes the path rearrangement model, the evidence generation model and the answer-level rearranger respectively, so that each module can work together with fixed parameters in the online reasoning phase, thereby improving the stability and auditability of the wind turbine generator fault diagnosis process.

[0107] (5) This invention has been validated on the WindTurbine-KG knowledge graph for wind turbine generator fault diagnosis. This graph is constructed from wind turbine generator equipment manuals and fault case records, containing 6558 entity nodes, 7476 relation edges, and 7485 sets of question-answer data. In the case of "What is the root cause of nacelle shaft track growth?", the system corrected the initial high-scoring but remedial answer "bolt re-tightening and torque verification" to "insufficient lubrication" which has continuous causal relationship support, and output the traceable path "nacelle shaft track growth → tooth surface pitting → insufficient lubrication" and natural language evidence. On the public dataset WebQSP, the Hit@1 and F1 of the complete method are 83.57% and 83.08%, respectively; the corresponding stage ablation results are shown in Table 4. The Hit@1 and F1 of the basic answer generation stage are 79.47% and 67.83%, respectively, indicating that consistency verification and backtracking correction can improve answer selection and enhance evidence traceability.

[0108] The above results are based on quantitative testing of public datasets and verification results from WindTurbine-KG cases. They are used to illustrate the technical process of candidate answer selection, path tracing, and evidence presentation, and are not presented as on-site deployment of wind turbine generators, conclusions of physical faults, or acceptance results by engineering experts. Actual engineering applications still require the combination of knowledge graph quality control, equipment status data, and review by domain experts. Attached Figure Description

[0109] Figure 1 This is the overall framework diagram of the present invention.

[0110] Figure 2 This is a flowchart of the consistency verification and backtracking correction process of the present invention. Detailed Implementation

[0111] The method described in this invention is geared towards fault diagnosis of wind turbine generator sets. Based on a knowledge graph for fault diagnosis of wind turbine generator sets, it organizes the entities of wind turbine generator set equipment, rotor, main shaft, bearing, gearbox and generator components, as well as fault phenomena, fault causes, detection methods, handling measures and fault impact diagnosis entities into a structured graph in the form of entity-relationship-entity. Then, it sequentially completes anchor point entity identification, candidate path retrieval, path rearrangement, evidence generation, answer aggregation, consistency verification and backtracking correction.

[0112] Experimental data included publicly available knowledge graph reasoning datasets ComplexWebQuestions, MetaQA, and WebQSP, as well as a self-built knowledge graph for wind turbine fault diagnosis, WindTurbine-KG. The publicly available datasets were used to test multi-hop knowledge graph reasoning capabilities; WindTurbine-KG, constructed from wind turbine equipment manuals and fault case records, was used to test fault localization, root cause analysis, evidence tracing, and answer correction processes.

[0113] WindTurbine-KG contains 6,558 entity nodes, 7,476 relation edges, and 7,485 sets of question-answer data, covering wind turbine equipment entities, rotor, main shaft, bearing, gearbox, and generator component entities, as well as entities for fault phenomena, fault causes, detection methods, handling measures, and fault impact diagnosis. It is used to demonstrate the reasoning path, natural language evidence, and backtracking correction results of this method in specific fault diagnosis problems.

[0114] The present invention adopts the following technical methods and implementation steps:

[0115] First, the system acquires information about wind turbine generator failures and identifies anchor entities, relational constraints, and diagnostic intents from the wind turbine generator failure diagnosis knowledge graph. Then, it performs constrained multi-hop path retrieval starting from the anchor entities. Subsequently, it fuses and rearranges the candidate paths using semantic and structural features, and converts high-quality paths into natural language evidence. Next, it forms candidate answers based on the path endpoints and aggregates supporting paths and evidence. Finally, it performs structured consistency and semantic consistency verification, and backtracks to select a credible answer from the candidate answer set when the current answer is inconsistent with the question intent or evidence chain.

[0116] The detailed steps are shown below. The formulas listed below repeatedly use the previously defined symbols.

[0117] (1) Basic Representation and Candidate Path Retrieval: First, a knowledge graph representation of wind turbine generator fault diagnosis is established, and candidate paths are retrieved based on this representation. In the retrieval stage, search depth, node expansion scale, and single relation type expansion constraints are introduced to control the number of candidate paths and reduce noise caused by high-connection components and high-frequency relations.

[0118] 1) Knowledge Representation and Path Definition. The knowledge graph for wind turbine generator fault diagnosis is denoted as a directed heterogeneous graph. .picture The entity types include wind turbine generator equipment entities; component entities include rotors, main shafts, bearings, gearboxes, and generators; diagnostic entities include fault phenomena, fault causes, detection methods, remedial measures, and fault effects. Relationship types include composition, association, cause, detection, remediation, and impact. Composition and association relationships are used for path connections between equipment, components, and fault phenomena. Cause, detection, remediation, and impact relationships correspond to fault cause, detection method, remedial measure, and fault impact diagnostic results, respectively. Facts in the graph are recorded in a triplet format of "head entity—relationship—tail entity".

[0119] The natural language wind turbine generator faults to be diagnosed are denoted as follows: The system first identifies one or more anchor entities, and then uses these anchor entities as the starting point of the path, expanding hop-by-hop along the directed relation edges in the knowledge graph. Any multi-hop reasoning path is represented as follows:

[0120] (1)

[0121] 2) Constrained Path Expansion. After completing the basic definition, the system performs a constrained breadth-first search starting from the set of anchor entities. The purpose of this process is to prioritize recalling paths that may support the answer, while avoiding path explosion caused by high-frequency relationships or highly connected nodes in complex graphs. The initial candidate path set is represented as follows:

[0122] (2)

[0123] in, This represents the knowledge graph for fault diagnosis of the same wind turbine generator set mentioned above; This indicates a constrained breadth-first search; Indicates from the question The set of anchor point entities identified in the graph, whose elements are graphs. Entity nodes in; Indicates the maximum search depth; This indicates the maximum number of candidate edges that can be retained for each node; This represents the maximum number of candidate edges that can be retained for each relation type under the same node; This represents the initial set of candidate paths obtained under the above constraints.

[0124] Each expansion consists of two steps. The first step is to sort the candidate edges according to their priority in equation (3) from highest to lowest, and retain at most [number missing] edges. The first step is to select candidate edges. The second step is to group them by relation type, ensuring that the number of edges retained for each relation type does not exceed a certain limit. The search depth reached Alternatively, stop expanding when the current layer no longer generates new paths.

[0125] , and All are pre-defined positive integers and satisfy the following conditions: Not greater than The constraints imposed by the three on the search process are... Pass it on to the next step.

[0126] During anchor entity recognition, the question text is first converted to lowercase and consecutive spaces are merged. Then, the longest exact match is performed on the entity name and alias in the knowledge graph. If multiple longest matches of the same length exist, all are retained, and entities with the same name are distinguished by unique node identifiers in the graph. If no exact match is obtained, 3-word, 2-word, and 1-word phrase matches are performed sequentially. If no anchor entity is still identified, an empty anchor set is returned.

[0127] Problem constraints are determined using executable dictionary and template rules. The system will then define the graph. The constraint dictionary consists of relation names, relation aliases, and common diagnostic question templates, mapping the prompt words in the question to relation signatures, expected hop counts, and expected endpoint roles. Specifically, "cause" and "root cause" are mapped to cause-type relations and fault cause-type endpoints; "detection" and "inspection" are mapped to detection-type relations and detection method-type endpoints; "handling," "repair," and "treatment" are mapped to handling-type relations and handling measure-type endpoints; and "impact" and "consequence" are mapped to impact-type relations and fault impact-type endpoints. The matched original prompt words constitute... The mapped relation type constitutes When the problem explicitly specifies one, two, or three hops, the corresponding value will be used as the expected number of hops; otherwise, the expected number of hops will be within a certain range. If no relationship suggestion is found, then... and If no endpoint role is identified, the endpoint check only requires... .

[0128] , and This is used to strike a balance between recall scope and computational overhead. The maximum number of inference hops required to cover the target task; Based on the average out-degree of knowledge graph nodes and available computing resources; Not greater than This is used to suppress high-frequency relationships from filling all candidate edges. Specific values ​​are shown in Table 1.

[0129] 3) Edge Priority Evaluation and Path Selection. During each hop expansion, the system does not retain all adjacent edges. Instead, it evaluates candidate edges based on question terms, target entities, relational hints, and semantic similarity, prioritizing relations more aligned with the question's intent. Candidate edge priority is expressed as follows:

[0130] (3)

[0131] Set as the complete set of terms for the normalization problem. Relationship hint term set. As an optional input to the path retrieval function; when not provided separately, the code uses the question normalization terms as relation hint terms. Semantic similarity is calculated using a frozen sentence vector encoder, which calculates the cosine similarity between the question text and the "relation text + target entity text". The relation term weights, target entity term weights, relation hint weights, and semantic similarity weights are configured as shown in Table 1.

[0132] After completing the multi-hop expansion, the system obtains the overall priority of the path based on the priority of each edge within the path, and retains the candidate paths with the highest global ranking:

[0133] (4)

[0134] (5)

[0135] The number of paths globally retained during the retrieval phase is determined by Limited, and record the search ranking starting from 1. If the number of candidate paths is insufficient, they will not be supplemented. See Table 1 for specific values.

[0136] (2) Path rearrangement and evidence generation

[0137] Candidate path retrieval focuses on recall and may still include paths that appear relevant but lack sufficient causal logic for the failure. To further filter high-quality evidence paths, the re-ranking stage utilizes both semantic matching features and structural statistical features, and transforms the retained structured paths into readable evidence.

[0138] 1) Semantic and structural feature fusion. For the retrieval candidate path set obtained by equation (5) The first in Candidate paths The system will display the problem text. and path text Inputting the same sentence into the vector encoder yields the semantic representations of both sentences:

[0139] (6)

[0140] A frozen all-MiniLM-L6-v2 sentence vector encoder with an output dimension of 384 was used; the encoder parameters were not updated during path reordering training and online inference. Starting with the first entity in the path, add each hop relationship and the target entity in sequence, separated by spaces.

[0141] (7)

[0142] (8)

[0143] The nine-dimensional structural features are all calculated directly from the problem and the path: the path hop count is taken from the number of relation edges, the initial ranking starts from 1, the text length is taken from the number of normalized terms, the two overlap values ​​are respectively counted as the number of common terms between the problem and the complete path text, and between the problem and the destination name, and the relation diversity is counted as the number of different relation types; the pattern node is judged according to the normalized node name pattern identifier registered in the knowledge graph construction stage, the path-level indicator checks each hop target node, and the destination indicator only checks the path endpoint.

[0144] Semantic and structural features are respectively processed in their respective branches for representation transformation, and then fused at the output layer to obtain the path relevance score.

[0145] (9)

[0146] The semantic branch input dimension is four times the frozen sentence vector dimension, that is... The input dimension of the structure branch is 9. and Both adopt a feature transformation structure of "fully connected layer - ReLU activation - Dropout", with a hidden dimension of 128 and a Dropout probability of 0.1; after concatenating the outputs of the two branches, the result is processed by parameters... and The linear output layer obtains a single rearranged score. During training, only the two branches and the output layer parameters are updated, using the AdamW optimizer. The learning rate and 5 training epochs are used to minimize the average loss of positive and negative path pairs within the same problem according to equation (22); the sentence vector encoder is kept frozen and the random seed is 42. Before preserving descending order Path formation .

[0147] 2) Path Linearization and Evidence Generation. The high-quality paths retained after rearrangement are still structured triple sequences, which are not convenient for manual review. Therefore, the system first decomposes the paths into ordered triples and adds structure tags to obtain linearized path text:

[0148] (10)

[0149] During linearization, the head entity, relation, and tail entity of each hop are written into the structured text in the actual hop order. The relation type and hop number in the path remain unchanged from the previous search results and are used for subsequent evidence generation and path verification.

[0150] Subsequently, the evidence generation model transcribes the linearized path into a natural language evidence description:

[0151] (11)

[0152] (12)

[0153] (3) Answer aggregation, consistency verification and backtracking correction

[0154] Different candidate paths may lead to the same answer, or they may lead to maintenance measures, failure consequences, or reasons for overly broad granularity. The system first aggregates the path support pointing to the same answer, and then combines the results of structured consistency and semantic verification to determine whether the current answer is reliable; if it is inconsistent with the question intent or the chain of evidence, a backtracking correction is performed.

[0155] 1) Answer aggregation. The system will rearrange and retain the set of candidate paths. By deduplicating the endpoint entities of each path, a set of candidate answers is obtained. Then, the supporting paths and evidence corresponding to each candidate answer are merged according to the path endpoint entity:

[0156] (13)

[0157] (14)

[0158] Before calculating equation (14), the number of supporting paths, the highest path rearrangement score, and the average path rearrangement score are respectively subjected to min-max normalization within the candidate answer set of the current question; when the maximum value of a certain item in all candidate answers is equal to the minimum value, that item is uniformly set to 1. The ranking of the best supporting path starts counting from 1 and uses its reciprocal in the fusion. The above processing makes the five input items of equation (14) within a comparable numerical range.

[0159] in, Indicates candidate answers The aggregate score; This represents the positive class matching probability output by the answer-level rearranger according to equation (25); , and These represent the number of supporting paths, the highest supporting path reorder score, and the average supporting path reorder score, respectively, within the candidate answer set for the same question after min-max normalization. If the maximum and minimum values ​​of a certain quantity are the same among all candidate answers, then its normalization result is uniformly set to 1. This indicates the ranking of the best support path, starting from 1. The reciprocal item indicates the reciprocal of this ranking, and its value is greater than 0 and does not exceed 1. to The weights are non-negative and sum to 1, with each of the five weights set to 0.2. The aggregated score is used in equation (15) to select the initial answer and forms the comprehensive verification ranking score in equation (20).

[0160] (15)

[0161] 2) Structured Consistency Verification. To avoid the answer being determined solely by local high-scoring paths, the system performs posterior checks on three aspects: evidence support strength, relation signature consistency, and path structure consistency. The three types of structured consistency signals are calculated as follows:

[0162] (16)

[0163] (17)

[0164] (18)

[0165] in, Pick . The path must start from the anchor entity, with each adjacent triplet being continuous from beginning to end, and must point from the anchor to the candidate answer along the configured directed relation. The path hop count must fall within the expected range of the aforementioned problem; Require And when the endpoint role has been identified, the role constraint is satisfied at the same time. When, equation (17) takes 1 and skips the relation signature constraint.

[0166] 3) Semantic Validation and Backtracking Correction. The structured consistency signal provides computable constraints, and the semantic validation model further determines whether there is consistency between the question, the current answer, and the evidence text. The system parses the raw output of the semantic validation model into structured validation results:

[0167] Before semantic verification, the system sorts the candidate answers in descending order of aggregate score and calculates the score difference between the top two candidate answers. When the score difference is less than the two-candidate arbitration threshold, the system fine-tunes the large language model by freezing the input parameters of the top two candidate answers, their supporting paths, comprehensive evidence, and three types of structured consistency values. Otherwise, only the first one is input. The input template contains the question text, candidate answer, supporting path, evidence support strength, relation signature consistency, path structure consistency, and comprehensive evidence in sequence, and requires the model to output only the two fields "best_answer" and "reason". The system parses the original output into the "support", "correction", or "undetermined" state as shown in Equation (19) based on the completeness of the fields, whether the answer belongs to the input candidate set, and whether the corresponding supporting path is not empty. When the fields are missing, the output answer does not belong to the candidate set, the corresponding supporting path is empty, or the parsing fails, it is recorded as "undetermined", and backtracking selection is performed by Equation (20) and Equation (21).

[0168] Semantic verification uses a fixed prompt template and deterministic decoding: temperature is set to 0, the maximum number of input lexical units is set to 4096, and the maximum number of newly added lexical units is set to 128. The first verification input is the current candidate answer and its supporting path, comprehensive evidence, and three types of structured consistency values; if the model proposes a corrected answer, the second verification only inputs the corrected answer and its corresponding supporting materials, and requires the output to be "supported" or "not determined". If any field is missing, the answer is out of bounds, the supporting path is empty, or the parsing fails at any stage, the model output is not used directly, but backtracking is performed according to Equations (20) and (21). The semantic verification model is not trained or updated in the online inference stage.

[0169] (19)

[0170] (20)

[0171] in, Represents the set of candidate answers to the same question. Internally, for aggregate scores The result obtained after performing exponential normalization; the hat symbol indicates the normalized quantity. In the summation, the first... 10 candidate answers; Indicates candidate answers The comprehensive verification ranking score, , , and represents the fusion weights for normalized aggregation score, evidence support strength, relation signature consistency, and path structure consistency, respectively; all are non-negative and sum to 1. In the actual calculation of exponential normalization, the maximum aggregation score for the current problem is subtracted from each exponential term to avoid numerical overflow; this processing does not change the normalization result. Equation (21) is used to perform backtracking selection when semantic verification does not directly determine the answer.

[0172] (twenty one)

[0173] This mechanism enables the system to backtrack from the set of candidate answers and select a more reliable diagnostic result when the initial answer does not conform to the question's intent or the chain of evidence is insufficient.

[0174] (4) Model training and online inference strategies

[0175] This invention employs a phased training strategy: the path rearrangement model, the evidence generation model, and the answer-level rearranger are trained separately, while the sentence vector encoder and the semantic verification large language model remain frozen during the online reasoning phase.

[0176] 1) Path Reordering Model Training. The path reordering model is used to learn the relative ordering relationship that "the correct path should precede the noisy path". For each question in the training set, the system labels positive and negative paths based on the standard answer entity and optimizes the path scores using pairwise ordering loss:

[0177] (twenty two)

[0178] The path reordering model was optimized using an adaptive moment estimation optimizer (AdamW) with decoupled weight decay. The learning rate, number of training epochs, sorting interval, and number of positive and negative paths extracted per problem were configured as shown in Table 1. The sentence vector encoder remained frozen, and only the parameters of the two-branch network and the output layer were updated.

[0179] 2) Training the evidence generation model. The evidence generation model is used to learn the mapping from structured paths to natural language evidence.

[0180] Training samples are constructed from problems with standard path annotations. Each standard path is converted into structured linear text according to Equation (10) and used as model input; each jump-head entity, normalized relation text, and tail entity are sequentially combined to form a statement, which serves as target evidence. Repeated input-target pairs are retained only once.

[0181] (twenty three)

[0182] 3) Training the answer-level rearranger. For candidate answers... Construct a ten-dimensional answer-level feature vector Its ten components are, in order, the number of overlapping terms between the question and the candidate answer, the overlap ratio, the number of terms in the question, the number of terms in the candidate answer, the number of supporting paths, the highest reorder score of the supporting path, the average reorder score of the supporting path, the best supporting path ranking, the last supporting path ranking, and whether there is a non-empty supporting path.

[0183] (twenty four)

[0184] (25)

[0185] The answer-level rearranger is trained using the binary cross-entropy loss shown in the following formula:

[0186] (26)

[0187] After offline training, the weight vector and bias of the answer-level logistic regression model are saved. During online inference, the parameters are loaded, and the correct answer prediction probability shown in Equation (25) is calculated based on the ten-dimensional feature vector of each candidate answer. This probability is used as the first term in Equation (14) for answer aggregation. The answer-level rearranger parameters are not updated during the online inference stage. The relevant fixed configuration and training parameters are summarized in Table 1.

[0188] Table 1 Main Fixed Configurations and Training Parameters

[0189]

[0190] Table 1 presents a set of reproducible configurations based on the current wind turbine generator fault diagnosis data. Quantities directly calculated from the problem, candidate paths, and knowledge graph are obtained in real-time according to their respective definitions; fixed configurations are determined on a validation set outside the training set, and trainable parameters are determined by minimizing equations (22), (23), and (26), respectively. When the wind turbine generator model, component range, or knowledge graph size changes, the fixed configuration is re-determined according to the same rules; training and experimental reproduction uniformly use a random seed of 42. The experimental dataset setup and training results are shown in Tables 2 and 3.

[0191] Table 2 Experimental Dataset Settings

[0192]

[0193] The public dataset is used to quantify multi-hop question-answering performance; WindTurbine-KG is used for case validation of wind turbine generator fault diagnosis. WindTurbine-KG is constructed from wind turbine generator equipment manuals and fault case records, containing 6558 entity nodes, 7476 relation edges, and 7485 sets of question-answer data. Unlike public benchmark datasets used for quantitative evaluation, WindTurbine-KG is primarily used for qualitative analysis and case studies, demonstrating reasoning paths, natural language evidence generation, and backtracking-based answer correction processes.

[0194] Based on the aforementioned datasets, embedding representation, graph structure reasoning, and large language model enhancement methods were selected as comparison objects. The Hit@1 and F1 scores for each method on different datasets are listed in Table 3. Hit@1 represents the proportion of the top-ranked predicted answer that matches the standard answer; F1 represents the harmonic mean of precision and recall calculated based on the predicted answer set and the standard answer set. Table 3 presents the results using a two-tiered header of "Dataset Name - Evaluation Metric," facilitating simultaneous comparison of the top-ranked answer hit rate and answer set matching degree of different models.

[0195] The comparison methods include EmbedKGQA, KV-MEM, TransferNet, UniKGQA, NSM, RAH, ReaRev, NuTrea, ChatKBQA, and DoG. Table 3 reports the Hit@1 and F1 results of this method and the comparison methods on multiple knowledge graph question answering datasets; unreported metrics are indicated by "—", and the numerical units in the table are %.

[0196] Table 3 Overall Comparison Results of the Models

[0197]

[0198] As shown in Table 3, the methods described above each have their advantages on different datasets, but their outputs typically focus on answer prediction or local path selection. In contrast, our proposed method achieves a Hit@1 score of 83.57 and an F1 score of 83.08 on the WebQSP dataset, maintains high performance on the MetaQA multi-hop task, and achieves superior F1 results on the CWQ dataset. This performance improvement is mainly attributed to the closed-loop mechanism constructed in this invention: "candidate path retrieval—semantic structure fusion and rearrangement—evidence generation—answer aggregation—consistency verification and backtracking correction." In the candidate path retrieval stage, irrelevant path expansion is suppressed through search depth, single-node expansion upper limit, and single-relationship type expansion upper limit. In the semantic structure fusion and rearrangement stage, the semantic matching relationship between the question and the path, as well as three types of structural information—path length, endpoint type, and relation diversity—are considered simultaneously to improve the ranking position of high-quality paths. In the path evidence generation stage, structured paths are transformed into natural language evidence, making the answer readable. In the answer aggregation and consistency verification stage, information from multiple supporting paths is further combined to perform structured and semantic verification of the initial answer, and backtracking correction is performed when the answer and evidence are inconsistent. The results demonstrate that candidate path constraint retrieval, semantic structure fusion and rearrangement, answer aggregation, and consistency verification can synergistically improve the stability of diagnostic result selection in complex multi-hop reasoning.

[0199] Table 4. Interim ablation results of the consistency verification mechanism (WebQSP, unit: %)

[0200]

[0201] The basic answer generation stage includes candidate path retrieval, path rearrangement, evidence generation, and answer aggregation, but does not perform subsequent consistency verification. The final stage, based on relation signature consistency and path structure consistency verification, further incorporates semantic verification, rejection, and backtracking correction mechanisms. Interim results show that the aforementioned consistency verification and backtracking correction processes can improve the selection of the first answer and increase the matching degree of the answer set.

[0202] Based on validation using publicly available datasets, to further verify the applicability of this method in wind turbine generator fault diagnosis scenarios, a root cause analysis of the wind turbine generator rotor and related components was selected. The input question was "What is the root cause of the nacelle shaft track growth?". After identifying the fault phenomenon entity, the system sequentially performed candidate path retrieval, path rearrangement, evidence generation, and answer verification. The initial high-scoring answer, "bolt retightening and torque verification," belonged to the treatment method category and was inconsistent with the root cause query intent. After consistency verification and backtracking, "insufficient lubrication" was determined, and the corresponding supporting path and natural language evidence were retained.

[0203] To compare the differences in evidence presentation and reasoning verifiability among different methods, Table 5 lists the results of EmbedKGQA, TransferNet, ReaRev, NuTrea, and our method on this case.

[0204] Table 5. Comparison of multi-model interpretability in wind turbine generator fault diagnosis cases.

[0205]

[0206] Table 5 shows that other methods can provide answers, relevant nodes, confidence scores, or structured subgraph paths, but they still lack explicit relational chains, natural language evidence, or consistency verification. This method provides both supporting paths and textual evidence, and corrects for inconsistencies between the answer type and the query intent.

[0207] To further illustrate the answer correction process, Table 6 lists the relational sequences and verification conclusions of several candidate answers. The system checks each candidate answer based on the query intent, relational type, path structure, and endpoint role, eliminating candidates whose handling measures and failure consequences are inconsistent with the root cause problem, and selecting a reliable answer from the remaining candidates.

[0208] Table 6 Analysis of Retrospective Amendments

[0209]

[0210] In this case, the initial top-ranked endpoint, "bolt retightening and torque verification," is a remedial measure and does not align with the root cause analysis intent; "insufficient lubrication" is supported by a continuous causal relationship, and its endpoint type is consistent with the analysis intent. Figure 1 The result was consistent with the established structure and passed semantic verification, thus being identified as the final answer. This outcome also preserves supporting pathways and natural language evidence, facilitating the tracing of the diagnostic conclusion's formation process.

Claims

1. An interpretable industrial fault diagnosis method based on knowledge graph multi-hop reasoning, characterized in that, Includes the following steps: (1) Obtain the wind turbine generator set fault problems to be diagnosed, and identify the anchor point entities and constraint relationships of the problems in the wind turbine generator set fault diagnosis knowledge graph; (2) Perform constrained multi-hop path retrieval starting from the anchor entity, expanding the candidate paths hop by hop; each node retains at most There are candidate edges, among which This indicates the upper limit for retaining candidate edges for a single node, set to 500, and forms a set of candidate inference paths; (3) The candidate path set is semantically and structurally fused and rearranged to select high-quality paths that conform to the fault propagation relationship of wind turbine generator sets and generate corresponding natural language evidence. (4) Generate a set of candidate answers based on high-quality paths, and aggregate the paths and evidence that support the same candidate answer to obtain a preliminary diagnostic answer; (5) Perform structured and semantic consistency verification on the preliminary diagnostic answer, and backtrack to select the final credible answer when the verification is inconsistent, and output the diagnostic results, supporting path and evidence description.

2. The method according to claim 1, characterized in that, The constrained multi-hop candidate path retrieval includes hop-by-hop expansion starting from anchor entities, and constrains the maximum number of inference hops, the expansion scale of a single node, and the number of single relation types to be retained during the expansion process. (1) Knowledge representation and path definition; the knowledge graph for fault diagnosis of wind turbine generator sets is denoted as a directed heterogeneous graph. ;picture The entity types include wind turbine generator equipment entities; component entities include rotors, main shafts, bearings, gearboxes, and generators; diagnostic entities include fault phenomena, fault causes, detection methods, remedial measures, and fault effects; relationship types include composition, association, cause, detection, remediation, and impact, where composition and association relationships are used for path connections between equipment, components, and fault phenomena, and cause, detection, remediation, and impact relationships correspond to fault cause, detection method, remedial measures, and fault impact diagnostic results, respectively; facts in the graph are recorded in the form of "head entity - relationship - tail entity" triples. The natural language wind turbine generator faults to be diagnosed are denoted as follows: The system first identifies one or more anchor entities, and uses these anchor entities as the starting point of the path, expanding hop-by-hop along the directed relation edges in the knowledge graph; any multi-hop reasoning path is represented as: (1)； in, This represents a path composed of entities and relations arranged in a logical order. This represents the number of relation edges contained in the path, i.e., the path hop count; for any Its value ranges from 1 to , No. Jump from the head entity ,relation Tail-end entity Composition, and the jump corresponds to the diagram. A fact triple in the middle; Represents the anchor point entity at the start of the path. Indicates the endpoint entity of the path; (2) Constrained path expansion; After completing the basic definition, the system performs a constrained breadth-first search starting from the set of anchor entities; the purpose of this process is to prioritize recalling paths that may support the answer, while avoiding path explosion caused by high-frequency relationships or highly connected nodes in complex graphs; the initial candidate path set is represented as: (2)； in, This represents the knowledge graph for fault diagnosis of the same wind turbine generator set mentioned above; This indicates a constrained breadth-first search; Indicates from the question The set of anchor point entities identified in the graph, whose elements are graphs. Entity nodes in; Indicates the maximum search depth; This indicates the maximum number of candidate edges that can be retained for each node; This represents the maximum number of candidate edges that can be retained for each relation type under the same node; This represents the initial set of candidate paths obtained under the above constraints; each expansion first retains the candidate edges according to the priority of equation (3). There are 10 candidate edges, and the number of edges retained for each relation type does not exceed 100. Search reached Or stop when no new paths are generated; , and All are pre-defined positive integers and satisfy the following conditions: Not greater than The constraints imposed by the three on the search process are... Passed to subsequent steps; Maximum search depth Based on the typical levels of the wind turbine generator fault propagation chain and the recall of candidate paths, the current value is 3; the number of candidate edges retained per node is... The number of nodes retained is determined by combining the average out-degree of the graph nodes, the recall rate of candidate paths, and the computational cost; currently, it is set to 500. within no more than Under the condition of limiting the expansion of high-frequency relationships, the current value is 200; when the wind turbine generator model, component range or knowledge graph size changes, the three are re-determined according to the same rules. Problem constraints are identified based on relation names, relation aliases, and diagnostic question prompts: cause or root cause corresponds to cause-type relations and fault cause-type endpoints; detection or inspection corresponds to detection-type relations and detection method-type endpoints; handling, repair, or treatment corresponds to handling-type relations and handling measure-type endpoints; impact or consequence corresponds to impact-type relations and fault impact-type endpoints; no relation signature constraints are applied when no relation prompt is identified; and only the consistency between the path endpoint and the candidate answer is checked when no endpoint role is identified. (3) Edge priority evaluation and path selection; During each hop expansion process, the system does not retain all adjacent edges, but evaluates candidate edges based on question terms, target entities, relational hints, and semantic similarity, prioritizing relations that are more consistent with the question intent; the priority of candidate edges is expressed as: (3)； in, This indicates the hop number of the current path, the hop number. Jump candidate edges are ; Indicate the problem The priority score of the single-hop candidate edge under the given conditions only evaluates the current hop and does not represent the score of the entire path; This represents the set of all terms after the problem is normalized. This represents a set of terms indicating the relationship between the question and the answer. This indicates text normalization and term segmentation. Indicates the number of elements in the set. This indicates that the entity name will be returned. To represent the intersection of sets, This indicates string concatenation. The sentence vector cosine similarity between the question text and the concatenated text of the relational text and the target entity name; , , and The non-negative weights control relation term matching, target entity term matching, relation prompt matching, and semantic similarity, respectively; the score is used for hop-by-hop candidate edge selection in equation (2) and as a hop-by-hop summation term in equation (4); All four weights are non-negative numbers, determined by combining the candidate path recall and ranking effects on the validation set; the configurations used are 1, 0.5, 2.5 and 1.0 respectively. After completing the multi-hop expansion, the system obtains the overall priority of the path based on the priority of each edge within the path, and retains the candidate paths with the highest global ranking: (4)； (5)； in, Indicate candidate path The overall retrieval priority is calculated from the score of each jump candidate edge by equation (3). to The scores are obtained by summing them up sequentially. When the path contains only one hop, the score of the path is equal to the score of the edge obtained by formula (3) for that hop. When the path contains two or more hops, the score of the path is the sum of the scores of each hop edge. This means retaining a specified number of elements based on their given scores, from highest to lowest. Indicates the number of globally reserved paths. Indicates from The most retained The set of candidate search paths obtained after searching for paths; the number of candidate paths is less than Preserve the entire path at the time; The number of paths to retain in the global filter is determined by the coverage of candidate paths and the cost of subsequent rearrangement, and the configuration used is 100.

3. The method according to claim 1, characterized in that, The process of semantic and structural fusion rearrangement of the candidate path set and generation of natural language evidence includes text encoding, structural feature extraction, dual-branch feature fusion and path linearization of the candidate reasoning paths to obtain the rearrangement score of the candidate reasoning paths and the corresponding fault diagnosis evidence description. (1) Semantic and structural feature fusion; for the retrieval candidate path set obtained by equation (5) The first in Candidate paths The system will display the problem text. and path text Inputting the same sentence into the vector encoder yields the semantic representations of both sentences: (6)； in, Indicate the problem semantic vector; Indicates the first Candidate paths semantic vectors, This represents the path text composed of entity names and relation names in the path, arranged in the path order. The sentence vector encoder represents the parameter freezing; both types of vectors are generated by the same sentence vector encoder and therefore have the same dimension. The sentence vector encoder uses the parameter-frozen all-MiniLM-L6-v2 with an output dimension of 384. Its parameters are not updated during path reordering training and online inference. (7)； in, Indicates the first The semantic interaction feature vectors of the candidate paths; each vector is concatenated along the feature dimension in the order listed in equation (7); This means taking the absolute value of the difference between corresponding elements of two vectors; This indicates element-wise multiplication; after concatenating the four parts, the dimension of this vector is four times the dimension of a single sentence vector. (8)； in, Indicates the candidate path number. Indicates the first Candidate paths The nine-dimensional structural eigenvectors; Indicates the number of hops in the path; This indicates the initial search ranking, starting from 1. Indicates the reciprocal of that ranking; Indicates the number of terms in the path text; The number of common terms between the question and the complete path text; The number of common terms representing the entity names of the problem and the path endpoint, where The endpoint entity of the path; Indicates the number of different relation types in the path; Indicates whether a pattern node exists in each hop target node of the path; The value is 1 if the path endpoint is a pattern node and 0 otherwise. The nine structural features in equation (8) form a structural feature vector in the order listed and serve as the input to the structural branch of equation (9). All nine components are calculated directly from the current question and candidate paths: the path hop count is the number of relation edges, the initial search ranking starts counting from 1, the reciprocal of the ranking is obtained from that ranking, the path text length and the overlap of the two items are counted according to normalized terms, and the relation diversity is counted according to different relation types; the path-level pattern node indicator is 1 when any target node matches the pattern identifier pre-registered in the knowledge graph, otherwise it is 0, and the endpoint pattern node indicator only checks the path endpoint; Semantic and structural features are respectively processed in their respective branches for representation transformation, and then fused at the output layer to obtain the path relevance score. (9)； in, and These represent the feature transformation functions for semantic branches and structural branches, respectively. and These represent the output vectors of the two branches, and the two output vectors are concatenated in the listed order. and These represent the weight vector and bias of the output layer, respectively. Take the transpose during calculation. Indicates the first The reordering score of each path; the trainable parameters mentioned above are determined by minimizing the pairwise ranking loss of equation (22), and the sentence vector encoder parameters remain frozen; this reordering score has a different meaning from the path priority in the retrieval stage of equation (4); according to Descending order retains at most The path is given, and the set of rearranged paths is obtained. And use it for candidate answer formation and path evidence aggregation in equation (13); Both semantic and structural branches adopt the "fully connected layer - ReLU activation - Dropout" structure; the hidden dimension, Dropout probability, and number of paths retained after rearrangement are determined by combining the validation set ranking effect and computational cost, and the configurations used are 128, 0.1, and 20, respectively. (2) Path linearization and evidence generation; the high-quality paths retained after rearrangement are still structured triple sequences. Therefore, the system first decomposes the path into ordered triples and adds structure labels to obtain the linearized path text: (10)； in, Indicates the candidate path number; Indicates the first The hop number within the path ranges from 1 to... ; , and The first one is formed in sequence. The jump's head entity, relation, and tail entity; Representing a path Structured linear text; "triples" The arrows and symbols are used to distinguish the jumps and their connection order. according to Concatenate the triplets in ascending order; Candidate paths are formed by hopping from anchor entities and combining the relational constraints identified by the current fault problem. The relational types and number of hops in the path are determined according to the specific fault problem. For any candidate path that is retained, the system linearizes each jump head entity, relation, and tail entity in sequence according to the actual jump order. Each jump relation is determined based on the actual directed connections of composition, association, cause, detection, handling, and impact relations in the knowledge graph. Subsequently, the evidence generation model transcribes the linearized path into a natural language evidence description: (11)； in, This indicates a sequence-to-sequence evidence generation model using the T5-small architecture; This represents the set of trainable parameters for the model, including encoder, decoder, and output layer parameters, and is determined by minimizing the generated negative log-likelihood loss of equation (23). This represents the structured linear path text obtained according to equation (10); Indicates by the first Natural language evidence text generated from the path; (12)； in, Indicates the first Comprehensive evidence corresponding to each path; This is an indicator of the validity of evidence, and its value is either 0 or 1. Represents structured linear path text; Represents natural language evidence text; when the generated text is not empty, contains the names of the first and last entities of the path, and the number of characters does not exceed [a certain threshold]. hour, Set the value to 1, and retain both types of text; otherwise, set the value to 0, and revert to structured linear path text. The maximum number of characters allowed for the validity of evidence is 300.

4. The method according to claim 1, characterized in that, The structured consistency verification and semantic consistency verification include verifying the evidence support strength of candidate answers, relation signature consistency, path structure consistency, and question-answer-evidence semantic consistency. (1) Answer aggregation; The system will rearrange and retain the set of candidate paths. By deduplicating the endpoint entities of each path, a set of candidate answers is obtained. Then, the supporting paths and evidence corresponding to each candidate answer are merged according to the path endpoint entity: (13)； in, This represents the set of candidate paths retained after rearrangement. Indicates the first Path, Representing a path The final entity, Indicates the first 10 candidate answers Indicates the endpoint entity is The set of supporting paths; Representing a path The comprehensive evidence formed according to formula (12), Indicates and The path in the equation corresponds to a comprehensive set of evidence; Equation (13) merges supporting paths and evidence according to the path endpoints, and the conditions after the vertical lines are used to limit the members of the set. The double vertical lines indicate the number of elements in the set; this equation is a set construction relationship and does not represent probability; Candidate answer set Depend on The endpoint entity is obtained by deduplication, so each candidate answer corresponds to at least one supporting path; (14)； in, Indicates candidate answers The aggregate score; This represents the positive class matching probability output by the answer-level rearranger according to equation (25); , and These represent the number of supporting paths, the highest supporting path reorder score, and the average supporting path reorder score, respectively, within the candidate answer set for the same question after min-max normalization. If the maximum and minimum values ​​of a certain quantity are the same among all candidate answers, then its normalization result is uniformly set to 1. This represents the ranking of the best support path, starting from 1. The reciprocal term in the formula represents the reciprocal of this ranking, and its value is greater than 0 and does not exceed 1. to The weights are non-negative and sum to 1. The comparison of the validation set adopts an equal weight setting, and all are set to 0.

2. The aggregated score is used to select the initial answer in equation (15) and forms the comprehensive validation ranking score in equation (20). (15)； Equation (15) is derived from the set of candidate answers. Selected from The largest candidate answer; The initial diagnostic answer obtained from this is used as the current answer for semantic verification in equation (19) and final answer selection in equation (21); when the aggregate scores are the same, the answer with priority is selected. The smaller candidate answer; if they are still the same, then keep the original candidate ranking. (2) Structured consistency verification; To avoid the answer being determined solely by local high-scoring paths, the system performs posterior checks on three aspects: evidence support strength, relation signature consistency, and path structure consistency; The three types of structured consistency signals are calculated as follows: (16)； (17)； (18)； in, Indicates candidate answers Strength of evidence; molecular accumulation The scores for each non-negative reordering path are calculated by summing the denominators. Scores for all non-negative multiple permutations of all paths; Used to truncate negative scores to 0. To prevent positive smoothing constants with denominators of 0; where, This represents the set of relation signatures obtained from the problem relation hint mapping. Indicates candidate answers The set of relation types contained in the supporting path. Representing a path No. The relationship of jumping, This indicates the number of hops in the path; Indicates consistency of relation signatures, when When not empty, take the number of elements in the intersection of the two sets and The ratio, when When the time is 1; where, , and These are used to check the path direction and continuity, the number of hops in the path, and the path's endpoint and its role. A value of 1 is returned if the condition is met, and 0 otherwise. This represents the structural matching value of a single support path; This represents the average value of all supporting path structure matching values; all three types of consistency results are used as inputs to equation (20); Direction and continuity indicators are based on the path starting from the anchor point entity, the continuity of two adjacent jump triples, and the relationship direction with the graph. The hop count indicator is set to 1 when the hop count falls within the expected range of the problem; the endpoint indicator is set to 1 when the endpoint equals the candidate answer and belongs to the entity type corresponding to the identified diagnostic intent (i.e., fault cause, detection method, remedial action, or fault impact); otherwise, it is set to 0. When the hop count is not explicitly stated in the problem, the expected range is 1 to 3. When the diagnostic intent is not identified, only the consistency between the endpoint and the candidate answer is checked. The evidence support smoothing constant is set to 1 × 10⁻. 8 This is only used to avoid a denominator of 0; (3) Semantic verification and backtracking correction; the structured consistency signal is used to provide computable constraints, and the semantic verification model further judges whether there is consistency between the question, the current answer and the evidence text; the system parses the original output of the semantic verification model into structured verification results: Before semantic verification, the system sorts the candidate answers in descending order of aggregate score and calculates the score difference between the top two candidate answers. When the score difference is less than the two-candidate arbitration threshold, the system fine-tunes the large language model by freezing the input parameters of the top two candidate answers, their supporting paths, comprehensive evidence, and three types of structured consistency values. Otherwise, only the first one is input. The input template contains the question text, candidate answer, supporting path, evidence support strength, relation signature consistency, path structure consistency, and comprehensive evidence in sequence. The system requires the model to output only the two fields "best_answer" and "reason". The system parses the original output into the "support", "correction", or "undetermined" state as shown in Equation (19) based on the field completeness, whether the answer belongs to the input candidate set, and whether the corresponding supporting path is not empty. When the field is missing, the output answer does not belong to the candidate set, the corresponding supporting path is empty, or the parsing fails, it is recorded as "undetermined", and backtracking selection is performed by Equation (20) and Equation (21). The two candidate arbitration thresholds are determined by combining the answer selection effect on the validation set, and the threshold used is 0.005; the semantic validation adopts deterministic decoding, the temperature is set to 0, the maximum number of input lexical units is set to 4096, and the maximum number of newly added lexical units is set to 128. (19)； in, This indicates that the system outputs the structured verification results generated by parsing based on the semantic verification model; Choose "Support", "Revise", or "Undetermined"; This represents the candidate answer selected by the semantic verification model, and can be empty in the "undecided" state; Natural language descriptions indicating the judgment made; (20)； in, Represents the set of candidate answers to the same question. Internally, for aggregate scores The result obtained after performing exponential normalization; the hat symbol indicates the normalized quantity. In the summation, the first... 10 candidate answers; Indicates candidate answers The comprehensive verification ranking score, , , and These represent the fusion weights for the normalized aggregation score, evidence support strength, relation signature consistency, and path structure consistency, respectively. They are all non-negative and their sum is 1. When actually calculating the index normalization, the maximum aggregation score under the current problem is subtracted from each index term to avoid numerical overflow. This process does not change the normalization result. Equation (21) is used to perform backtracking selection when semantic verification does not directly determine the answer; The weights for the four validation rankings are all non-negative and sum to 1, determined by combining the validation set results; the configuration used is an equal weight setting, with each weight set to 0.

25. (21)； in, This represents the corrected answer proposed by semantic verification. This indicates the status obtained after performing a second verification on the corrected answer; when the semantic verification status is "supported", Get the initial answer When the semantic verification status is "corrected", the system first checks... and A second confirmation will only be performed after the above checks have passed; if the checks fail or the confirmation fails, an order will be issued. "Undetermined"; when For "correction" and When "support", Pick In other cases, from Select the comprehensive verification ranking score Largest candidate answer; confirmed Then, output the answer along with its corresponding set of supporting paths and a comprehensive set of evidence.

5. The method according to claim 1, characterized in that, The model training and online inference strategy includes phased training of the path reordering model, evidence generation model and answer-level reorderer, and keeping the sentence vector encoder and semantic verification large language model frozen during the online inference stage. (1) Path reordering model training; The path reordering model is used to learn the relative ordering relationship that "the correct path should be placed before the noisy path"; For each question in the training set, the system labels the positive and negative paths according to the standard answer entities, and uses pairwise ordering loss to optimize the path score: (22)； in, Represents the positive path under the same problem With negative path Pair sorting loss; and These represent the rearrangement scores for positive and negative paths, respectively. Indicates the pre-defined sorting interval; The value is the larger of 0 and the value in parentheses; a positive path is a path where the endpoint entity belongs to the standard answer set; a negative path is a path where the endpoint entity does not belong to the standard answer set; negative paths are divided into high-ranking difficult negative paths and other medium-difficulty negative paths according to their search ranking, and are paired with positive paths to form training sample pairs; the average loss of all positive and negative paths for the same question is calculated using formula (22), and the average loss is used for one parameter update; the gradient optimization method is used to minimize the loss shown in formula (22), and the model parameters determined after training are used for online path rearrangement; The sorting interval and the number of positive and negative paths extracted were determined by combining the path discrimination effect on the validation set and the training cost. The configuration used was 0.1, with a maximum of 4 positive paths, 8 difficult negative paths and 8 medium-difficulty negative paths per problem. The AdamW optimizer was used with a learning rate of 1×10⁻³, 5 training rounds, and a random seed of 42. (2) Training the evidence generation model; the evidence generation model is used to learn the mapping from structured paths to natural language evidence; training samples are constructed from questions with standard path annotations: each standard path generates input text according to equation (10), and each jump-head entity, normalized relation text, and tail entity are sequentially combined into a statement as target evidence; repeated "input-target" pairs are only retained once; the training set and validation set are divided according to the question number, and all path samples of the same question are guaranteed to enter the same data subset; the input is denoted as The target word sequence is denoted as The generation loss is expressed as: (23)； in, Represents the target evidence word sequence; Indicates the total number of target lexical units; Indicates the current position; Indicates the first One target word element; Indicates the first to the second Given a known target word, when This indicates an initial condition that includes only the sequence start marker; The parameter is The conditional probabilities given by the evidence generation model; The sum of the negative log-likelihoods at each position is represented; the loss shown in equation (23) is minimized using the sequence-to-sequence generation model to determine the parameters of the evidence generation model; the character count threshold for evidence validity checking. The upper limit of the word count in the decoding stage is used for post-generation checks and constraints during the generation process, respectively. The evidence generation model uses T5-small; the relevant configuration is determined by combining the convergence of training loss and the quality of evidence on the validation set, with a learning rate of 5×10⁻. 5 Training for 3 rounds, batch size of 8, beam search width of 4, maximum number of input lexical units of 512, maximum number of new lexical units of 300, and random seed of 42; (3) Answer-level rearranger training; the answer-level rearranger is used to determine whether candidate answers are consistent with the standard answer; for candidate answers Construct a ten-dimensional answer-level feature vector The ten components are, in order: the number of overlapping terms between the question and the candidate answer, the overlap ratio, the number of terms in the question, the number of terms in the candidate answer, the number of supporting paths, the highest re-ranking score of the supporting path, the average re-ranking score of the supporting path, the best supporting path ranking, the last supporting path ranking, and whether there is a non-empty supporting path. (24)； in, Indicates candidate answers The corresponding ten-dimensional answer-level feature vector; the ten components represent, in turn, the number of normalized terms overlapping between the question and the candidate answer, the overlap ratio, the number of terms in the question, the number of terms in the candidate answer, the number of supporting paths, the highest re-ranking score of the supporting path, the average re-ranking score of the supporting path, the best supporting path ranking, the last supporting path ranking, and the supporting path existence identifier. exist The value is 1 if the condition is met, and 0 otherwise; the ten components form a column vector in the order listed. (25)； in, and These represent the weight vector and bias of the answer-level logistic regression model, respectively. Represents a logical function; A binary variable representing the correctness of candidate answers. This indicates that the candidate answer is consistent with the standard answer. This indicates inconsistency; Indicates that given a feature vector Under the given conditions, the predicted probability that a candidate answer is the correct answer. The predicted probability; The answer-level rearranger is trained using the binary cross-entropy loss shown in the following formula: (26)； in, Indicates the first Ten-dimensional answer-level feature vectors of training samples. This represents the predicted probability that the sample belongs to the positive class. Represents a true binary label; This represents the training objective of the answer-level rearranger. The total number of training samples is represented; it is determined by minimizing equation (26). and After training, the parameters are fixed, and the positive class probability of the candidate answer is calculated according to Equation (25) during the online inference stage. In the implementation, a class-balanced logistic regression classifier is used, with a maximum number of iterations of 1000 and a random seed of 42. Save formula (25) after offline training is completed and The parameters are loaded during online inference, and calculations are performed based on the ten-dimensional feature vector of each candidate answer. And this probability is used as the first term of equation (14) to participate in answer aggregation; the answer-level rearranger parameters are no longer updated during the online reasoning stage; The trainable parameters of the path rearrangement model, the evidence generation model, and the answer-level rearranger are determined by minimizing the training objectives of equations (22), (23), and (26), respectively.

Images