Intention-aware neural-symbolic graph retrieval enhanced fault diagnosis method

Abstract

The application relates to the technical field of fault diagnosis, and specifically discloses an intention-aware neural symbolic graph retrieval enhanced fault diagnosis method, which comprises the following steps: converting unstructured operation and maintenance text into a weighted probability causal graph embedding fault mechanism based on a physical information graph construction protocol; identifying the dominant intention of user query and generating an intention modulation matrix through a four-quadrant intention routing mechanism to dynamically adjust the graph edge weight; fusing physical rules and semantic similarity by adopting a double-flow neural symbolic adaptive pruning algorithm to screen key reasoning paths; and constructing reasoning context based on the screened paths and inputting a large language model to generate a diagnosis report. Through the cooperation of intention awareness and physical logic constraints, the application realizes accurate retrieval and reasoning of fault causal chains, significantly improves the reliability and explainability of industrial fault diagnosis, and is suitable for man-machine collaborative diagnosis in safety-critical scenarios.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically to a fault diagnosis method enhanced by intention-aware neural symbol graph retrieval. Background Technology

[0002] Currently, modern industrial equipment is gradually developing towards larger scale, greater complexity, and higher intelligence. System resilience and human-machine collaboration have become the core objectives of the digital transformation of the manufacturing industry. Under this paradigm, intelligent operation and maintenance systems have been upgraded from traditional passive data recording tools to cognitive enhancement partners supporting the entire lifecycle management of equipment. Industrial sites accumulate massive amounts of data through sensor networks and manual recording. Among this data, unstructured operation and maintenance texts, which carry implicit experience and equipment physical mechanisms, contain key knowledge such as fault propagation patterns and maintenance decision-making logic. These are core resources for solving unseen faults and improving diagnostic reliability.

[0003] Traditional fault diagnosis technologies are mainly divided into two categories: one is a physical model-based system that relies on precise mathematical modeling to describe the operating rules of equipment. However, when faced with the nonlinear, time-varying characteristics and cross-system coupled faults of modern equipment, the modeling difficulty is extremely high and the adaptability is poor. The other is a data-driven deep learning method that achieves fault classification by mining the statistical features of sensor time-series data. However, its "black box" nature leads to a lack of interpretability in the diagnostic results, making it impossible to associate the physical mechanisms in unstructured text and difficult to gain trust in safety-critical scenarios.

[0004] To overcome these limitations, knowledge graph technology has been introduced into the field of industrial fault diagnosis. By extracting triples to construct structured knowledge networks, it attempts to achieve logically traceable reasoning. However, existing industrial knowledge graphs are mostly static and flat structures, simply representing the relationships between entities without explicitly quantifying the physical strength of the relationships and the probability of fault propagation. Furthermore, they struggle to distinguish the logical priority of causal and structural relationships. When such graphs are applied to fault diagnosis, the high density of structural relationships easily creates topological noise, masking the sparse but crucial causal chains of faults and causing the reasoning direction to deviate from the core requirements.

[0005] While neurosymmetric AI technology has achieved a fusion of data-driven and knowledge-driven approaches, existing systems are mostly intent-agnostic, failing to consider the engineering-oriented nature of user queries in industrial operations and maintenance scenarios. Fault diagnosis corresponds to reverse causal reasoning, impact analysis to forward evolutionary reasoning, and spatial localization to structural reasoning; different intents require different reasoning paths. Retrieval systems lacking intent guidance cannot dynamically adjust the reasoning direction, making it difficult to meet the precise requirements of heterogeneous industrial tasks.

[0006] Therefore, how to provide an efficient fault diagnosis reasoning method that can embed fault mechanisms, perceive user intent, and integrate physical rules and semantic features is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed to provide an intention-aware neural symbol graph retrieval-enhanced fault diagnosis method that overcomes or at least partially solves the above problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] An intention-aware neural symbol graph retrieval-enhanced fault diagnosis method specifically includes the following steps: S1. Obtain unstructured operation and maintenance data of the target device, clean the unstructured operation and maintenance data, use the sliding window algorithm to slice the cleaned unstructured operation and maintenance data to generate a set of text slices, and use the physical information graph construction protocol to convert the text slices in the text slice set into a weighted probability causal graph.

[0010] S2. Receive the user's natural language query, identify the dominant intent of the natural language query through the four-quadrant intent routing mechanism, and generate an intent modulation matrix.

[0011] S3. Dynamically adjust the edge weights of the weighted probabilistic causal graph based on the dominant intent of the natural language query, and lock the topological traversal direction that matches the dominant intent of the natural language query.

[0012] S4. The dual-stream neural symbol adaptive pruning algorithm is adopted to calculate the comprehensive transmission rate of the topological traversal direction that matches the dominant intent. It integrates physical rule constraints and semantic similarity matching, combines the comprehensive transmission rate and calculates the cumulative energy of each inference path according to the path energy accumulation formula, sets the cumulative energy threshold, and filters the inference paths in the weighted probability causal graph. Paths below the cumulative energy threshold are filtered out, and key paths above the cumulative energy threshold are retained.

[0013] S5. Construct a reasoning context based on the critical path above the energy threshold, input the reasoning context into the large language model, generate a preliminary diagnostic report, perform a logical consistency check on the preliminary diagnostic report, and check whether the causal relationship in the preliminary diagnostic report is consistent with the causal relationship in the weighted probability causal graph. If they are inconsistent, automatically supplement the energy calculation process of the critical path and return to S4; if they are consistent, generate a logically consistent and traceable diagnostic report.

[0014] Preferably, the specific process in S1 of converting text slices in the text slice set into a weighted probability causal graph based on the physical information graph construction protocol includes: A fault evolution ontology set is defined through a domain constraint protocol. The concept of text slices is classified into the corresponding ontology categories in the ontology set using a mapping function. The ontology categories include at least fault modes, fault phenomena, physical components, maintenance measures, and environmental factors.

[0015] Based on the semantics of entity relationships in text slices, the extracted relationships are classified into types, including causal relationship types and structural relationship types. A strong causal relationship subset is constructed based on the causal relationship type, and a weak structural relationship subset is constructed based on the structural relationship type. The strong causal relationship subset is used to represent the main chain of fault propagation and the correlation between fault phenomena and fault mechanisms, while the weak structural relationship subset is used to represent the compositional membership relationship or spatial positional relationship between equipment components.

[0016] The extracted relations correspond to candidate relation edges. For each extracted candidate relation edge, a confidence quantification protocol is executed. The confidence quantification protocol includes: constructing relation evidence units corresponding to the candidate relation edges. The relation evidence units include at least source text fragments, head entities, relation types, and tail entities; outputting the original confidence scores of the candidate relation edges from the large language model based on the relation evidence units; and then assigning corresponding relation type adjustment coefficients according to the relation types to which the candidate relation edges belong, so as to jointly quantify the credibility and relation semantic strength of the candidate relation edges.

[0017] The text slices in the text slice set are transformed into a weighted probabilistic causal graph, where the weighted probabilistic causal graph is formally defined as a weighted directed graph. , V For a collection of entities, E Let be a set of directed edges. W Let be the set of static physical weights for the edges.

[0018] Preferably, the formula for calculating static physical weights is:

[0019] in, The original confidence scores of candidate relation edges generated by the large language model based on relation evidence units are given. The value range is [0,1]; This is a relation type adjustment coefficient; different relation types correspond to different values, and causal relations correspond to... The value is greater than the value corresponding to the structural class relation. Values; For the balance parameters, and satisfying This is used to balance the contribution ratio of the original confidence score and the relation type adjustment coefficient in the static physical weight.

[0020] Preferably, the sliding window in S1 is sliced, with the window length set to 500-800 characters and the overlap length set to 50-100 characters to ensure the integrity of the causal relationship.

[0021] Preferably, the implementation process of the four-quadrant intent routing mechanism in S2 includes: We employ a neural-symbolic cascade reasoning strategy. First, we calculate the posterior probability distribution of the query belonging to four types of intent using a large language model. Then, we combine this with a domain rule base based on regular expressions for hard constraint verification.

[0022] The four types of intents include fault diagnosis intent, impact analysis intent, spatial positioning intent, and general query intent.

[0023] Preferably, the fault diagnosis intent corresponds to the reverse causal reasoning requirement. The reverse causal reasoning requirement refers to taking the fault phenomenon or abnormal performance as the starting point and tracing back along the reverse causal relationship to locate the fault cause or source of the fault phenomenon or abnormal performance.

[0024] Impact analysis intent corresponds to the need for forward evolutionary reasoning. Forward evolutionary reasoning refers to taking the cause of the failure, the failure event, or the abnormal state as the starting point and expanding along the positive direction of the causal relationship to analyze the possible subsequent impacts or evolutionary results.

[0025] Spatial positioning intent corresponds to structural reasoning requirements. Structural reasoning requirements refer to traversing the compositional and hierarchical relationships or spatial positional relationships between equipment components in order to determine the installation location, hierarchical level, or associated structure of the target component.

[0026] The general query intent corresponds to the unbiased semantic retrieval requirement, which means that the causal direction bias or structural direction bias is not preset, and the retrieval is based on the semantic relevance between the natural language query and text slices, entity nodes or reasoning paths.

[0027] Preferably, the formula for calculating the overall conduction rate of the dual-stream neural symbol adaptive pruning algorithm in S4 is as follows:

[0028] in, For overall conductivity, The neural symbol balance coefficient, and satisfies 0 < <1, The static physical weights of the edges. The dynamic gain coefficients in the intended modulation matrix, To query the semantic cosine similarity with the path, This is the Sigmoid activation function.

[0029] Preferably, the formula for calculating the cumulative path energy in S5 is:

[0030] in, For the first The cumulative energy of the path, This is the path length attenuation factor. For the first The out-degree of the step node.

[0031] Preferably, before the diagnostic report is generated in S5, a reasoning path visualization step is also included for a white-box presentation of the diagnostic process.

[0032] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an intention-aware neural symbol graph retrieval enhanced fault diagnosis method. The beneficial effects of the above technical solutions provided by the embodiments of the present invention include at least the following: 1) Achieve refined quantitative characterization of fault mechanisms: Through the construction protocol of physical information graphs, qualitative fault descriptions in unstructured texts are transformed into weighted causal relationships with confidence levels, clearly distinguishing the logical priority of causal relationships and structural relationships. This solves the problem that traditional static graphs cannot express the probability of fault propagation, making the graphs closer to the physical degradation laws of equipment. 2) Achieve precise alignment between intent and reasoning path: The four-quadrant intent routing mechanism can accurately identify the user's engineering-oriented intent, dynamically adjust the graph traversal strategy through the intent modulation matrix, avoid indiscriminate wandering of general retrieval, ensure that the focus is on causal chain backtracking during fault diagnosis, on consequence evolution during impact analysis, and on structural decomposition during spatial positioning, and significantly improve the adaptability of heterogeneous tasks. 3) Improve the credibility and interpretability of diagnostic results: The dual-stream neural symbolic pruning algorithm integrates physical rules and semantic features to effectively filter high-density structural noise and ensure the logical rigor of the reasoning path. At the same time, through visualization of the reasoning process and verification of logical consistency, the diagnostic results are traceable and verifiable, which solves the pain point that traditional black box models are difficult to trust in industrial scenarios and provides technical support for human-machine collaborative diagnosis in safety-critical scenarios. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0034] Figure 1 This is a schematic diagram of the overall process provided in the embodiments of the present invention; Figure 2 This is a knowledge graph of fault information for a special vehicle provided in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the performance of three ablation test indicators provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the radar performance of the ablation experiment provided in this embodiment of the invention in four types of tasks. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] This invention discloses an intent-aware neural symbol graph retrieval-enhanced fault diagnosis method, which specifically includes the following steps: S1. Obtain unstructured operation and maintenance data of the target device, clean the unstructured operation and maintenance data, use the sliding window algorithm to slice the cleaned unstructured operation and maintenance data to generate a set of text slices, and use the physical information graph construction protocol to convert the text slices in the text slice set into a weighted probability causal graph.

[0037] S2. Receive the user's natural language query, identify the dominant intent of the natural language query through the four-quadrant intent routing mechanism, and generate an intent modulation matrix.

[0038] S3. Dynamically adjust the edge weights of the weighted probabilistic causal graph based on the dominant intent of the natural language query, and lock the topological traversal direction that matches the dominant intent of the natural language query.

[0039] S4. The dual-stream neural symbol adaptive pruning algorithm is adopted to calculate the comprehensive transmission rate of the topological traversal direction that matches the dominant intent. It integrates physical rule constraints and semantic similarity matching, combines the comprehensive transmission rate and calculates the cumulative energy of each inference path according to the path energy accumulation formula, sets the cumulative energy threshold, and filters the inference paths in the weighted probability causal graph. Paths below the cumulative energy threshold are filtered out, and key paths above the cumulative energy threshold are retained.

[0040] S5. Construct a reasoning context based on the critical path above the energy threshold, input the reasoning context into the large language model, generate a preliminary diagnostic report, perform a logical consistency check on the preliminary diagnostic report, and check whether the causal relationship in the preliminary diagnostic report is consistent with the causal relationship in the weighted probability causal graph. If they are inconsistent, automatically supplement the energy calculation process of the critical path and return to S4; if they are consistent, generate a logically consistent and traceable diagnostic report.

[0041] To achieve the above objectives, the present invention adopts the following detailed technical solution: An intent-aware neural symbol graph retrieval-enhanced fault diagnosis method is designed, which integrates the logical interpretability of knowledge graphs, the physical constraints of neural symbol reasoning, and the efficiency of RAG technology. Specifically, it includes the following: Step 1: Constructing Unstructured O&M Data Preprocessing and Weighted Probabilistic Causal Graph. Obtain unstructured O&M data for the target equipment, including equipment maintenance manuals, fault case libraries, and maintenance logs. Clean the data by removing garbled characters, meaningless symbols, and duplicate text using regular expressions. Employ a sliding window algorithm to slice the cleaned data, setting the window length to 500-800 characters and the overlap length to 50-100 characters to ensure the integrity of fault causal relationships is preserved, generating a set of text slices.

[0042] The protocol based on physical information graphs transforms text slices into weighted probabilistic causal graphs. The specific process is as follows: 1) Define a set of fault evolution ontology, including five core ontology categories: fault mode, fault phenomenon, physical component, maintenance measure, and environmental factor. Use mapping functions to accurately map the concepts in the text slices to the corresponding ontology categories to avoid entity confusion. 2) A hybrid extraction strategy combining large language models and rule templates is adopted. By matching causal sentences such as "A causes B", "A triggers B", and "B is caused by A" through regular expressions, and combining the contextual understanding capabilities of large language models, entity pairs and relationships are extracted from text slices to construct an initial graph. 3) Relation type classification, relation subset construction, and static physical weight calculation: Based on the entity relation semantics extracted in the previous step, the relations are classified into types. The relation types include causal relation types and structural relation types. The causal relation type includes at least CAUSES, LEADS_TO, MANIFESTS_AS, and SYMPTOM relations, which are used to characterize the mapping relationship between fault causes, fault propagation, state evolution, and fault modes and observable phenomena. The structural relation type includes at least PART_OF and LOCATED_IN relations, which are used to characterize the compositional membership relationship and spatial location relationship between physical components.

[0043] Strong causal subsets are constructed based on causal relationship types, and weak structural subsets are constructed based on structural relationship types. The strong causal subsets are used to characterize the main chain of fault propagation and the correlation between fault phenomena and fault mechanisms, while the weak structural subsets are used to characterize the compositional hierarchy or spatial location relationships between equipment components, thereby distinguishing the logical priority of different relationships in subsequent graph traversal.

[0044] The confidence quantification protocol in this invention refers to a set of quantification rules that convert the textual evidence credibility and semantic strength of candidate relation edges into static physical weights. Specifically, a relation evidence unit is constructed for each candidate relation edge, and the relation evidence unit includes at least a source text fragment, a head entity, a relation type, and a tail entity; the relation evidence unit is input into a large language model, which outputs the original confidence score of the candidate relation edge; then, a corresponding relation type adjustment coefficient is assigned according to the relation type to which the candidate relation edge belongs, so as to jointly quantize the candidate relation edge.

[0045] The formula for calculating static physical weights is:

[0046] in, The original confidence score of the candidate relation edges generated by the large language model based on the relation evidence unit has a value range of [0,1]. This is the adjustment coefficient for the relation type, corresponding to causal relations. The value is greater than the value corresponding to the structural class relation. Values; To balance the parameters, a value of 0.7 can be used in this embodiment to prioritize the proportion of physical mechanism relationships in the graph edge weights, ultimately resulting in a weighted directed graph. This completes the explicit coding of the fault mechanism.

[0047] Step 2: Implement a four-quadrant intent routing mechanism. After receiving a user's natural language query, the dominant intent is identified through a neural-symbolic cascade reasoning strategy, generating an intent modulation matrix. The specific process is as follows: 1) Input user queries into a pre-trained large language model and calculate the posterior probability distribution of the query belonging to four intent categories: fault diagnosis, impact analysis, spatial localization, and general query. ,in ; 2) Construct a domain rule base and match intent feature words using regular expressions: fault diagnosis intent matches keywords such as "cause", "why", and "root cause"; impact analysis intent matches keywords such as "consequence", "impact", and "cause"; spatial positioning intent matches keywords such as "location", "installation", and "composition"; and general query intent has no specific feature words. 3) Combining the posterior probability with the rule base verification results, select the intent with the highest probability that conforms to the rule constraints as the dominant intent. If there are probabilities that are tied or rules that conflict, the intent confirmation is completed by querying the user to supplement the context.

[0048] In this invention, different dominant intentions correspond to different reasoning needs and topological traversal directions. Specifically, the fault diagnosis intention corresponds to the need for reverse causal reasoning, that is, starting from the fault phenomenon or abnormal manifestation, tracing back along the causal relationship edge to locate the source of the fault; the impact analysis intention corresponds to the need for forward evolutionary reasoning, that is, starting from the fault cause, fault event, or abnormal state, expanding along the causal relationship edge in the forward direction to analyze its subsequent impact; the spatial positioning intention corresponds to the need for structural reasoning, that is, traversing along the component composition relationship and spatial position relationship to determine the installation location, hierarchical level, or adjacent structure of the target object; and the general query intention corresponds to the need for unbiased semantic retrieval, that is, without pre-setting the causal relationship direction or structural relationship direction, but retrieving according to the semantic relevance between the natural language query and candidate text slices, entity nodes, or reasoning paths.

[0049] Based on the different needs mentioned above, in the subsequent step 3, different dynamic gain coefficients are assigned to different types of relation edges through the intent modulation matrix, so that the fault diagnosis intent prioritizes the causal backtracking path, the impact analysis intent prioritizes the consequence evolution path, the spatial positioning intent prioritizes the structural path, and the general query intent maintains a relatively balanced search preference.

[0050] Step 3: Determine the topological traversal direction for dominant intent matching. Generate the intent modulation matrix based on the dominant intent. Each row of the matrix corresponds to an intent, each column corresponds to three types of relations, and the elements represent the physical connectivity of the relations. Specific parameters are shown in Table 1. Table 1: Attack Rates for Different Types of Intents

[0051] The edge weights of the weighted probabilistic causal graph are dynamically adjusted using this matrix to obtain the adjusted edge weights. Lock the topology traversal direction that matches the intent.

[0052] Step 4: Execute the dual-stream neural symbol adaptive pruning algorithm. Based on the adjusted dynamic weight graph, the dual-stream neural symbol adaptive pruning algorithm is used to filter key inference paths. The specific process is as follows: 1) Starting from the corresponding entity node in the query, initialize the path queue, set the maximum path length to 4 hops, and the initial energy of the path. .

[0053] 2) Traverse each path in the path queue and calculate the overall transmission rate of the current edge. The formula is:

[0054] in, The neural symbol balance coefficient, To query the cosine similarity between the vector and the path text description vector, This is the Sigmoid activation function.

[0055] 3) Update the cumulative energy of each path according to the path energy accumulation formula, which is:

[0056] in, This is the path length attenuation factor. For the first The out-degree of the step node.

[0057] 4) Set an energy threshold, filter out paths with accumulated energy below the threshold, and retain the Top-K high-energy paths as key inference paths to complete topology noise filtering.

[0058] Step 5: Diagnostic Context Construction and Report Generation. Based on the selected key reasoning paths, a structured reasoning context is constructed, including descriptions of entity relationships corresponding to the paths, original text fragments, and weighted fusion values. This context is then input into the large language model to generate a preliminary diagnostic report.

[0059] Before the report is generated, a visualization step of the reasoning process is added: key information such as intent routing results, dynamic weight adjustment process, and energy changes of key causal chains are displayed through charts, realizing a white-box presentation of the diagnostic process, which makes it easier for operation and maintenance personnel to verify the reasoning logic.

[0060] Finally, the preliminary diagnostic report is checked for logical consistency: whether the causal relationship in the test report is consistent with the relationship in the weighted probability causal graph. If there is an undefined causal relationship in the graph, the energy calculation process of the relevant path is automatically supplemented. If the energy calculation result is lower than the threshold, it is indicated that the causal relationship has low credibility, ensuring that the report is logically rigorous and traceable.

[0061] This invention discloses an intent-aware neural symbol graph retrieval-enhanced fault diagnosis method through a specific embodiment, such as... Figure 1 As shown, taking the diesel engine system and electromechanical transmission system of a certain type of special vehicle as the application objects, and combining their unstructured operation and maintenance data to achieve fault diagnosis, the specific implementation steps are as follows: 1. Construction of weighted probability causal graph The core of the process is to transform unstructured text into a weighted graph that embeds fault mechanisms. The specific implementation is as follows: 1) Fault Evolution Ontology Definition: Five core ontologies are defined, including fault modes, fault phenomena, physical components, maintenance measures, and environmental factors. Concepts in text slices are accurately classified through mapping functions. For example, "bearing overheating" is mapped to the "fault phenomenon" ontology.

[0062] 2) Entity and relation extraction: A hybrid strategy of Qwen-max large language model and rule templates is adopted. The rule templates match causal sentences and extract entity pairs and relations from the slices. For example, the entity pairs (vibration, seal wear) and relation "CAUSES" are extracted from "vibration causes seal wear".

[0063] 3) Relationship Subset Construction and Static Physical Weight Calculation: The extracted entity relations are categorized according to relational semantics. Causal relation types include CAUSES, LEADS_TO, MANIFESTS_AS, and SYMPTOM relations; structural relation types include PART_OF and LOCATED_IN relations. Strong causal relation subsets are constructed based on causal relation types, and weak structural relation subsets are constructed based on structural relation types.

[0064] In this embodiment, the strong causal subset is used to represent the main chain of fault propagation and the correlation between fault phenomena and fault mechanisms. For example, relationships such as "vibration causes seal wear" and "incomplete combustion manifests as black smoke" can be classified into the strong causal subset. The weak structural subset is used to represent the compositional membership or spatial positional relationship between equipment components. For example, relationships such as "impeller belongs to pump body" and "injector is installed in cylinder head" can be classified into the weak structural subset.

[0065] A fault evolution ontology set is defined through a domain constraint protocol. The concept of text slices is classified into the corresponding ontology categories in the ontology set using a mapping function. The ontology categories include at least fault modes, fault phenomena, physical components, maintenance measures, and environmental factors.

[0066] Based on the semantics of entity relationships in text slices, the extracted relationships are classified into types. The extracted relationships correspond to candidate relationship edges. The relationship types include causal relationship types and structural relationship types. A strong causal relationship subset is constructed based on the causal relationship type, and a weak structural relationship subset is constructed based on the structural relationship type. The strong causal relationship subset is used to represent the fault propagation main chain and the association between fault phenomena and fault mechanisms, while the weak structural relationship subset is used to represent the compositional membership relationship or spatial position relationship between equipment components.

[0067] A confidence quantification protocol is used to calculate the static physical weight of each edge. The specific process of the confidence quantification protocol is as follows: First, a relation evidence unit is constructed for each candidate relation edge. The relation evidence unit includes at least the source text fragment, head entity, relation type, and tail entity. Second, the relation evidence unit is input into a large language model, which outputs the original confidence score of the candidate relation edge. Secondly, assign corresponding relation type adjustment coefficients based on the relation type to which the candidate relation edges belong. Finally, based on the relation type to which the candidate relation edges belong, corresponding relation type adjustment coefficients are assigned to jointly quantify the credibility and semantic strength of the candidate relation edges, thus adjusting the original confidence scores. Adjustment coefficient for relation type The static physical weights of each candidate relation edge are calculated using the static physical weight formula.

[0068] For example, regarding the text fragment "coolant leakage will inevitably lead to a pressure drop," since this relationship is of the causal type, a higher relationship type adjustment coefficient is used to calculate its static physical weights, and the original confidence level generated by the large model is... Then the edge weight is calculated as follows:

[0069] For the text fragment "pump body includes impeller", since this relationship belongs to the structural relationship type, a relationship type adjustment coefficient lower than that for causal relationships is used to calculate its static physical weight. The original confidence score generated by the large model is... The weights are:

[0070] The final result is a weighted directed graph. , V For a collection of entities, E Let be a set of directed edges. W Let be the set of static physical weights for the edges.

[0071] like Figure 2 As shown, it contains 1897 entities and 2420 edges, completing the explicit encoding of the fault mechanism.

[0072] 2. Implementation of the four-quadrant intent routing mechanism In this embodiment, the intent recognition module is implemented through large-scale model probability calculation and rule base verification. The specific process is as follows: 1) Construction of the intent feature word rule base: Fault diagnosis intent keyword set K1={cause, root cause, why, lead to}, impact analysis intent K2={consequence, impact, will cause, trigger}, spatial positioning intent K3={location, composition, installation, belong to}, general query intent has no specific keywords.

[0073] 2) Posterior probability calculation: Given the user query input to Qwen-Max, the output is the posterior probability of the four types of intents. For example, the probability distribution of the query "What is the root cause of white smoke from a diesel engine?" is as follows: .

[0074] 3) Intent Recognition: Combining probability and rule base, this query contains root cause keywords, and the fault diagnosis intent has the highest probability, thus it is determined to be the dominant intent; if the query is "What are the components of a centrifugal pump?", the probability distribution... The text contains keywords related to components, which is interpreted as an intention for spatial positioning.

[0075] 4) Intent Modulation Matrix Generation: Based on the judgment result, the corresponding conduction rate parameter is called. For example, under the intent of fault diagnosis, In the CAUSES / LEADS_TO relation, the connectivity is set to 3.0, and in the PART_OF / LOCATED_IN relation, it is set to 0.1. The edge weights are dynamically adjusted. .

[0076] 3. Execution of the Two-Stream Neural Symbolic Adaptive Pruning Algorithm In this embodiment, the pruning algorithm parameters are set as follows: maximum path length of 4 hops, initial energy... Neural symbol balance coefficient Path length attenuation factor With an energy threshold of 0.5 and Top-K=10, the specific execution is as follows: 1) Path initialization: Starting from the entity node corresponding to the query, for example, querying the entity "emitting white smoke" corresponding to "diesel engine emitting white smoke", initialize the path queue, which contains all edges that start or end with that node; 2) Overall conductivity calculation: Taking the path "coolant leakage → water ingress into combustion chamber → white smoke" as an example, the conductivity of the path "coolant leakage → water ingress into combustion chamber" is calculated as follows: Fault diagnosis intent Query semantic similarity with path text description Sigmoid activation function ,but:

[0077] 3) Path energy update: Assuming the energy of the previous hop... The current node's "water inlet in combustion chamber" output degree ,but:

[0078] 4) Noise filtering: Filter paths with energy below 0.5, retain the top-10 high-energy paths, and complete the screening of key causal chains.

[0079] 4. Diagnostic context construction and diagnostic report generation 1) Context construction: Extract text slices, entity relationship descriptions, and weight fusion values ​​corresponding to the Top-10 key paths to form a structured context, such as "coolant leakage leads to water entering the combustion chamber, which in turn causes white smoke" and "cylinder head gasket leakage will cause coolant to enter the combustion chamber"; 2) Visualization of the reasoning process: The intention routing results, dynamic weight adjustment values, and critical path energy change curves are displayed through charts, achieving a white-box presentation; 3) Logical consistency check: Check whether the causal relationship in the generated report exists in the graph. For example, if the relationship "coolant leakage → white smoke" in the report exists in the graph and the energy meets the standard, then the report shows "insufficient engine oil → white smoke". If the report shows "insufficient engine oil → white smoke", then a logical conflict is determined. 4) Report generation: After verification, Qwen-Max is called to generate a structured report, which includes the root cause of the failure, the propagation mechanism, and troubleshooting suggestions. If there is a logical conflict, it will automatically fall back to the extraction mode, extract the original sentence from the slice as the answer and mark it as "based on the original text".

[0080] 5. Experimental Verification and Performance Analysis To verify the effectiveness of this invention, a test set containing 200 question-answer pairs was constructed, including 60 diagnostic pairs, 60 impact analysis pairs, 40 localization pairs, and 40 general question-answer pairs. NaiveRAG and PathRAG were selected as benchmarks for comparison. To evaluate the adaptability and generalization ability of the method of this invention on different general large language model bases, this embodiment selected three representative large language models as underlying bases for testing, specifically: GPT-4o, Qwen-max, and Gemini-2.5-flash. Among them, GPT-4o is a high-intelligence flagship model provided by OpenAI, supporting text and image input; Qwen-max is a high-performance model in the Qwen series, suitable for complex, multi-step tasks; and Gemini-2.5-flash is a general-purpose model provided by Google that balances price and performance with low latency.

[0081] By selecting underlying models from different sources and with different performance focuses as experimental bases, the adaptability, stability, and generalization ability of the proposed method to different large language models were verified. It should be noted that the intent-aware and two-stream neural symbol graph retrieval enhancement method proposed in this invention is a general retrieval enhancement generation framework and does not depend on the network architecture of the specific large language model mentioned above. Any large language model with natural language understanding, context processing, and text generation capabilities can be applied to this invention. The results of testing on the above three model bases and the comparison of core performance indicators are shown in Table 2.

[0082] Table 2: Comparison of Core Performance Indicators of Different Methods

[0083] As shown in the table above, the root cause detection rate, diagnostic accuracy, and comprehensive diagnostic score of this invention are significantly better than the baseline on all model bases. Among them, the root cause detection rate of the GPT-4o base is improved by 7.4% and the diagnostic accuracy is improved by 9.8% compared with PathRAG, which proves the synergistic effect of intention perception and neural symbol pruning.

[0084] like Figure 3 As shown, ablation experiments further validated the necessity of the core module (based on the GPT-4o base). When the intent recognition module was removed, the system root cause detection rate dropped significantly by 11.9%, and the overall diagnostic score decreased by nearly 10%. This result indicates that the intent module is crucial for ensuring causal chain recall. When the neural symbol pruning module was removed, the system diagnostic accuracy plummeted from 79.1% to 64.9%, a decrease of 14.2%. This result confirms that the pruning module can effectively filter topological noise.

[0085] To further explore the differences in the adaptation of intent routing mechanisms to different industrial sub-tasks, such as Figure 4 As shown, performance radar charts for three types of models were plotted across four tasks: fault diagnosis, impact analysis, spatial localization, and general question answering. The radar chart of the complete model exhibits a full and balanced shape, indicating its stable generalization ability across various heterogeneous tasks. However, the performance curve of the variant without intent awareness does not show a uniform contraction characteristic but rather exhibits a significant topological bias, with a particularly pronounced performance drop in fault diagnosis and impact analysis tasks. This phenomenon suggests that without the intention module's directional weighting of causal edges, the random walk algorithm is easily captured by structural edges, making it difficult to penetrate high-density structural noise to uncover sparse fault causal chains, leading to a significant performance decline in causal reasoning tasks. In contrast, the performance curve of the variant without a pruning module shows a uniform contraction characteristic across all dimensions, indicating that noise filtering is a universal requirement for improving the accuracy of all types of tasks, and the logical purification role of the neural symbol pruning module is indispensable in various industrial scenarios.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault diagnosis method enhanced by intent-aware neural symbol graph retrieval, characterized in that, Includes the following steps: S1. Obtain unstructured operation and maintenance data of the target device, clean the unstructured operation and maintenance data, use the sliding window algorithm to slice the cleaned unstructured operation and maintenance data to generate a set of text slices, and convert the text slices in the set of text slices into a weighted probability causal graph based on the physical information graph construction protocol. S2. Receive the user's natural language query, identify the dominant intent of the natural language query through the four-quadrant intent routing mechanism, and generate an intent modulation matrix; S3. Dynamically adjust the edge weights of the weighted probability causal graph according to the dominant intent of the natural language query, and lock the topological traversal direction that matches the dominant intent of the natural language query. S4. The dual-stream neural symbol adaptive pruning algorithm is adopted to calculate the comprehensive transmission rate of the topological traversal direction that matches the dominant intent. It integrates physical rule constraints and semantic similarity matching, combines the comprehensive transmission rate and calculates the cumulative energy of each inference path according to the path energy accumulation formula, sets the cumulative energy threshold, and filters the inference paths in the weighted probability causal graph. Paths below the cumulative energy threshold are filtered out, and key paths above the cumulative energy threshold are retained. S5. Construct a reasoning context based on the critical path above the energy threshold, input the reasoning context into the large language model, generate a preliminary diagnostic report, perform a logical consistency check on the preliminary diagnostic report, and check whether the causal relationship in the preliminary diagnostic report is consistent with the causal relationship in the weighted probability causal graph. If they are inconsistent, automatically supplement the energy calculation process of the critical path and return to S4; if they are consistent, generate a logically consistent and traceable diagnostic report.

2. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 1, characterized in that, The specific process of converting text slices in a text slice set into a weighted probabilistic causal graph based on the physical information graph construction protocol in S1 includes: A fault evolution ontology set is defined through a domain constraint protocol. The concept of text slices is classified into the corresponding ontology category in the ontology set using a mapping function. The ontology category includes at least fault mode, fault phenomenon, physical component, maintenance measure and environmental factor. Based on the semantics of entity relationships in text slices, the extracted relationships are classified into types, including causal relationship types and structural relationship types. A strong causal relationship subset is constructed based on the causal relationship types, and a weak structural relationship subset is constructed based on the structural relationship types. The strong causal relationship subset is used to characterize the fault propagation main chain and the association between fault phenomena and fault mechanisms, and the weak structural relationship subset is used to characterize the compositional membership relationship or spatial position relationship between equipment components. The extracted relations correspond to candidate relation edges. For each extracted candidate relation edge, a confidence quantification protocol is executed. The confidence quantification protocol includes: constructing a relation evidence unit corresponding to the candidate relation edge, wherein the relation evidence unit includes at least a source text fragment, a head entity, a relation type, and a tail entity; outputting the original confidence score of the candidate relation edge from the large language model based on the relation evidence unit; and then assigning a corresponding relation type adjustment coefficient according to the relation type to which the candidate relation edge belongs, so as to jointly quantify the confidence level and relation semantic strength of the candidate relation edge. The text slices in the text slice set are transformed into a weighted probabilistic causal graph, where the weighted probabilistic causal graph is formally defined as a weighted directed graph. , V For a collection of entities, E Let be a set of directed edges. W Let be the set of static physical weights for the edges.

3. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 2, characterized in that, The formula for calculating the static physical weight is: in, The original confidence scores of candidate relation edges generated by the large language model based on relation evidence units are given. The value range is [0,1]; This is a relation type adjustment coefficient; different relation types correspond to different values, and causal relations correspond to... The value is greater than the value corresponding to the structural class relation. Values; For the balance parameters, and satisfying This is used to balance the contribution ratio of the original confidence score and the relation type adjustment coefficient in the static physical weight.

4. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 1, characterized in that, In S1, the sliding window uses a sliced ​​processing method, with the window length set to 500~800 characters and the overlap length set to 50~100 characters to ensure the integrity of the causal relationship.

5. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 1, characterized in that, The implementation process of the four-quadrant intent routing mechanism in S2 includes: The neural-symbolic cascade reasoning strategy is adopted. First, the posterior probability distribution of the query belonging to four types of intent is calculated through a large language model, and then hard constraint verification is performed by combining the domain rule base based on regular expressions. The four types of intents include fault diagnosis intent, impact analysis intent, spatial positioning intent, and general query intent.

6. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 5, characterized in that, The fault diagnosis intent corresponds to the reverse causal reasoning requirement, which means taking the fault phenomenon or abnormal performance as the starting node and tracing back along the reverse causal relationship to locate the fault cause or source of the fault phenomenon or abnormal performance. The impact analysis intent corresponds to the forward evolution reasoning requirement, which refers to taking the cause of the failure, the failure event, or the abnormal state as the starting node and expanding along the positive direction of the causal relationship to analyze the possible subsequent impacts or evolutionary results. The spatial positioning intent corresponds to the structural reasoning requirement, which refers to traversing along the compositional hierarchy or spatial positional relationship between equipment components to determine the installation location, hierarchical level, or associated structure of the target component. The general query intent corresponds to the unbiased semantic retrieval requirement, which means that the retrieval is performed based on the semantic relevance between the natural language query and text slices, entity nodes or reasoning paths without pre-setting causal or structural bias.

7. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 5, characterized in that, The formula for calculating the overall conduction rate in the dual-stream neural symbol adaptive pruning algorithm in S4 is: in, For overall conductivity, The neural symbol balance coefficient, and satisfies 0 < <1, The static physical weights of the edges. The dynamic gain coefficients in the intended modulation matrix, To query the semantic cosine similarity with the path, This is the Sigmoid activation function.

8. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 1, characterized in that, The formula for calculating the cumulative path energy in S5 is: in, For the first The cumulative energy of the path, This is the path length attenuation factor. For the first The out-degree of the step node.

9. The intention-aware neural symbol graph retrieval-enhanced fault diagnosis method according to claim 1, characterized in that, Before generating the diagnostic report, S5 also includes a reasoning path visualization step, which is used for a white-box presentation of the diagnostic process.

Images