An aerial intelligent question-answering system and method based on graph-module collaborative enhancement

Abstract

This application belongs to the field of interdisciplinary application technology of artificial intelligence and aviation, and specifically relates to an aviation intelligent question-answering system and method with graph-model collaborative enhancement. It generates factual question-answering samples, process question-answering samples, constraint judgment samples, counterfactual samples, and safety constraint samples based on a knowledge graph. Semantic expansion is then performed using a large language model, and the training set is enhanced by using graph relationship constraints, time validity constraints, and evidence consistency constraints. During the training phase, a dual-channel collaborative training method with question encoding and evidence encoding channels is employed. During the update phase, incremental updates of the shadow graph, simulated question-answering verification, incremental adapter learning, and version rollback mechanisms are used. During the answering phase, the system outputs the answer text, graph path, source fragment, time validity description, confidence level description, and risk warning. This solution can improve the professional reliability, real-time knowledge update capability, long-tail scenario coverage, and result interpretability of aviation question answering.

Description

Technical Field

[0001] This application belongs to the field of interdisciplinary application technology of artificial intelligence and aviation, and specifically relates to an aviation intelligent question answering system and method with graph-model collaborative enhancement. Background Technology

[0002] Aviation knowledge is characterized by its dispersed sources, complex terminology, high time sensitivity, and strict safety constraints. Question-and-answer systems designed for aviation operations, maintenance support, airworthiness compliance analysis, aeronautical information interpretation, and anomaly handling not only need to understand technical terminology, Chinese and English abbreviations, aircraft type differences, and procedural logic, but also need to consider the real-time impact of NOTAMs, weather changes, airport status, flight status, and fleet configuration differences when providing answers.

[0003] Existing aviation question-answering solutions can be broadly categorized into three types. The first type, centered on rule bases or keyword templates, can achieve rapid retrieval within a fixed question range, but struggles to adapt to the diverse expressions of questions, numerous long-tail scenarios, and frequently changing knowledge in the aviation environment. The second type, based on domain knowledge graphs, completes question answering through entity recognition, relation extraction, and graph retrieval. While providing some structured reasoning capabilities, it often relies on text-based triples, making it difficult to incorporate drawings, tables, voice recordings, and real-time operational data. Furthermore, it lacks sufficient support for complex procedural questions, constraint judgment questions, and cross-source evidence conflict issues. The third type directly applies general or domain-specific large language models to aviation question answering. While possessing strong semantic understanding and expressive capabilities, it is prone to issues such as answer illusions, untraceable evidence, the infiltration of time-lapsed information, and unstable judgments regarding aviation constraints.

[0004] Existing technologies suffer from at least the following shortcomings: First, the lack of unified modeling for the time validity, source reliability, evidence location, and version differences of aviation knowledge makes it difficult for the system to keep answers synchronized when operational notices, weather, and fault records change rapidly. Second, the scarcity of labeled samples for long-tail aviation scenarios makes it difficult for general question-and-answer training sets to cover specialized issues such as maintenance diagnosis, special weather operations, anomaly handling, and airworthiness restrictions. Third, existing knowledge graph question-and-answer systems and large language model question-and-answer systems are mostly simple concatenations, and have not yet formed a system centered around "knowledge graph generation samples, model expansion samples, knowledge graph re-verification samples, and dual-channel..." Fourth, existing systems typically lack a closed-loop mechanism for "training with consistent evidence" when outputting answers, making it difficult to meet the requirements of interpretability and traceability in aviation scenarios. Fifth, when real-time data is updated, most systems can only reconstruct the graph or retrain the entire dataset, resulting in high update costs and a lack of shadow verification and rollback mechanisms. Sixth, when source evidence conflicts or is insufficient, existing systems lack conflict resolution and rejection mechanisms based on validity period, source reliability, and strictness of restrictions, making it easy to output unstable conclusions in high-risk scenarios.

[0005] Therefore, there is an urgent need to provide a new intelligent question-answering solution for the aviation field, which can construct a multimodal knowledge graph with unified identification, time validity attributes and source fingerprints for aviation professional data, use graph-modal complementary data augmentation to alleviate the problem of long-tail sample scarcity, achieve verifiable, traceable and rollbackable question-answer output under the condition of dynamic knowledge updates, and automatically carry out risk control when there is evidence conflict or insufficient evidence. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a system and method for aviation intelligent question answering that enhances the collaboration between the aviation knowledge graph and the large language model. The solution is based on the following main lines: "generating samples with graph constraints, expanding samples with the model, training with evidence consistency, incremental updates triggered by events, conflict resolution and rejection control, and dual-dimensional evidence output". It transforms the simple superposition of aviation knowledge graph and large language model into a collaborative system with complementary functions and mutual verification.

[0007] According to a first aspect of this application, a graph-model collaborative enhancement aviation intelligent question answering system is provided, including a data access and governance module, a multimodal aviation knowledge graph construction module, a graph-model complementary data enhancement module, a joint training module, a dynamic knowledge update module, and an interpretable question answering engine.

[0008] The data access and governance module is used to access aviation operation regulations, aircraft manufacturer technical data, maintenance manuals, minimum equipment lists, configuration deviation lists, fault and handling records, operation notices, airport operation information, weather messages, flight status information, voice recordings, drawings, images and video data, and to perform format unification, entity mapping, noise removal, version identification, desensitization and timestamp alignment on data from different sources.

[0009] The multimodal aviation knowledge graph construction module is used to build an aviation knowledge graph covering aircraft type, system, component, failure mode, alarm, process steps, constraints, airport, runway, weather, flight, operational events, and handling suggestions. Unlike existing graphs that only record entities and relationships, this application adds unified identifiers, valid time intervals, source fingerprints, evidence locations, and confidence fields to graph nodes and relationships, enabling subsequent retrieval and reasoning to constrain the freshness of answers, source credibility, and citation location.

[0010] The graph-model complementary data augmentation module is used to generate an enhanced training set. This module first extracts templates from entity relationship pairs, process chains, constraint chains, and causal chains in the graph to form seed question-answer pairs for different question types, such as fact-finding, process handling, constraint judgment, comparative analysis, and dynamic situations. Then, it calls a large language model to perform terminology replacement, expression rewriting, colloquialization, cross-language transformation, and counterfactual and interference sample generation on the seed question-answer pairs. Finally, it filters the generated samples based on graph constraints, aviation rule constraints, valid time intervals, and evidence consistency constraints, retaining training samples that are logically closed, have identifiable sources, and do not conflict with constraints, while adding violating samples to the counterexample set. This alleviates the problem of scarce annotations in long-tail aviation scenarios and enhances the ability to refuse answers.

[0011] The joint training module is used to perform dual-channel collaborative training on the aviation question-answering model. The first channel is the question encoding channel, which performs semantic encoding on natural language questions; the second channel is the evidence encoding channel, which performs structured encoding on atlas subgraphs, source paragraphs, drawing location information, table items, and real-time data fragments. The two channels are aligned through cross-attention or gating fusion mechanisms, and jointly optimize the answer generation loss, evidence alignment loss, time validity constraint loss, and safe rejection loss. In this way, the model not only focuses on language similarity when generating answers, but also on whether the answer is supported by both the atlas path and the original source fragment.

[0012] The dynamic knowledge update module is used to update the shadow graph when changes occur in operational notices, weather messages, ADS-B status, airport operational status, maintenance work cards, or fault records. Simulated questions and answers are automatically generated around the change points on the shadow graph, and lightweight adapter incremental learning and validation are performed. When the validation results meet consistency, accuracy, and safety thresholds, the shadow graph and the updated model are switched to the official version; otherwise, a rollback is performed. This avoids the spread of errors caused by directly overwriting the official knowledge base.

[0013] An interpretable question-answering engine is used to execute multi-route answers based on question type and controls answer output through evidence consistency scoring. For fact-finding questions, graph retrieval and evidence concatenation results are output first; for process handling and constraint judgment questions, reasoning based on graph paths and rule clauses is executed first; for complex questions with insufficient evidence but which can be generated under constrained conditions, the question-answering model is invoked to generate output, along with source fragments, time validity descriptions, confidence levels, and risk warnings; for questions with conflicting evidence or high security risks, a rejection or supplementary confirmation prompt is output. The candidate evidence consistency score is calculated based on at least three of the following: graph structure matching degree, source reliability, time validity, and real-time data consistency; when multiple candidate pieces of evidence conflict, evidence that is still valid, has higher source reliability, and stricter constraints is prioritized.

[0014] According to a second aspect of this application, a graph-model collaborative enhancement method for intelligent question answering in aviation is provided, comprising the following steps:

[0015] S1. Collect structured data, semi-structured data, unstructured data and multimodal data in the aviation field, and perform deduplication, standardization, unified identifier mapping, timestamp alignment, timeliness attribute annotation and evidence location indexing on the data;

[0016] S2 extracts entities, relationships, attributes, constraints, process chains, and valid time intervals from the processed data to construct a multimodal aviation knowledge graph with unified identifiers, source fingerprints, and confidence levels;

[0017] S3 generates seed question-answer pairs based on the graph template, and then uses the large language model to generate rewritten samples, counterfactual samples, comparison samples, distractor samples and cross-language samples. An enhanced training set is formed through graph rules, time constraints and evidence consistency constraints, while samples that violate the constraints are written into the counterexample set.

[0018] S4. Input the question representation and the graph subgraph evidence representation into the joint training module for collaborative training to obtain the aviation domain question answering model;

[0019] S5, after receiving user questions, first classifies and routes the questions, then recalls the corresponding graph subgraphs, rule clauses, source paragraphs and real-time data, calculates the consistency score of candidate evidence, and generates an answer when the score meets the threshold, and outputs supplementary confirmation prompts, risk prompts or rejection results when the score does not meet the threshold.

[0020] S6 outputs the question-and-answer results, including the answer text, graph path, source fragment, time validity description, and confidence level description.

[0021] Preferably, in step S3, safety constraint samples are generated around operational restrictions, maintenance release, anomaly handling, and airworthiness constraints, and extended samples that violate the restriction relationships or process sequences are written into the counterexample set for subsequent rejection training.

[0022] Preferably, before step S5, a dynamic knowledge update process is also included: after detecting changes in operation notices, weather, flight status, or maintenance status, the information is first updated in the shadow map, and lightweight adapter incremental learning is performed on the question-answering model; only when both the simulated question-answering verification and the security verification pass are the formal map and the formal model replaced.

[0023] Compared with the prior art, this application has at least the following beneficial effects:

[0024] First, by introducing a unified identifier, a valid time interval, a source fingerprint, and a confidence level field into the aviation knowledge graph, answers can be jointly constrained during the retrieval and generation stages, reducing the probability of referencing outdated, invalid, or low-reliability knowledge.

[0025] Second, by using a graph-model complementary data enhancement loop of "graph generation samples, model expansion samples, graph re-verification samples, and writing violation samples into the counterexample set", the sample coverage of long-tail questions, variant questions, and mixed Chinese and English questions is improved, while the system's ability to refuse to answer high-risk questions is also enhanced.

[0026] Third, by co-training the question coding channel and the evidence coding channel, an explicit correspondence is established between the generated answer and the structured evidence, and the final output is controlled by the evidence consistency score, thereby enhancing the interpretability and verifiability of the answer.

[0027] Fourth, through shadow graphs, simulated question-and-answer verification, and rollback mechanisms, the system can absorb dynamic knowledge without full retraining and control the risk of erroneous updates entering the formal question-and-answer process.

[0028] Fifth, by using problem classification routing, conflict resolution rules, and security rejection constraints, the system can output supplementary confirmations, risk warnings, or rejection results in high-risk aviation scenarios, thereby improving application reliability. Attached Figure Description

[0029] Figure 1 This is the overall architecture diagram of the graph-model collaborative enhancement aviation intelligent question-answering system of this application;

[0030] Figure 2 This is a flowchart illustrating the data enhancement process for complementary graph models in this application.

[0031] Figure 3 This is a flowchart illustrating the dynamic knowledge update and rollback process for this application.

[0032] Figure 4 This is a flowchart of the interpretability questions and evidence output for this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are only some, not all, of the embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0034] The first aspect of this application provides a graph-model collaborative enhancement aviation intelligent question-answering system, such as... Figure 1 As shown, it can be deployed on airport operation support platforms, aircraft maintenance assistance platforms, aeronautical information analysis platforms, airworthiness compliance analysis platforms, or aviation enterprise knowledge service platforms. The system accesses data including, but not limited to: AMM, IPC, FCOM, fault isolation manuals, aircraft configuration lists, maintenance work cards, operational bulletins, airport operation databases, weather messages, ADS-B data, alarm logs, component drawings, and duty recordings. After unified access, each data source undergoes text parsing, table structuring, image OCR / layout understanding, speech recognition, and time alignment processing before entering the knowledge graph construction process.

[0035] In one embodiment, the multimodal aviation knowledge graph construction module includes the following process: First, an aviation-domain named entity recognition model is used to extract entities from the text. These entities include at least aircraft type, system, component, fault code, maintenance action, constraints, airport, runway, weather phenomenon, flight, and operational event. Then, a relationship extraction model and rule templates are used to extract relationships such as "component-belongs to-system," "fault-corresponds to-handling steps," "aircraft type-limited by-operating conditions," "airport-has-runway parameters," and "weather-affects-operating status." For drawings and diagrams, the system establishes a mapping between image objects and graph entities based on component numbers, area labels, and structural connections. For voice text, the system extracts event entities, instruction entities, and temporal sequence information. Finally, a unified identifier, source fingerprint, evidence location, effective time, expiration time, and confidence level are added to each graph relationship to form a multimodal graph that can be used for temporal reasoning and evidence localization.

[0036] In one embodiment, the graph-model complementary data augmentation process is as follows: Figure 2As shown, the system extracts representative path templates from the graph, such as "fault code → fault mode → handling action," "aircraft type → constraints → operational conclusion," "weather phenomenon → impact relationship → airport operational status," and "procedure step A → preceding → procedure step B," and generates seed question-and-answer pairs accordingly. Subsequently, the large language model expands the seed question-and-answer pairs around different question styles, generating abbreviation-based questions, colloquial questions, abbreviated questions, mixed Chinese and English questions, comparative questions, and counterfactual and distractor questions. To prevent the expanded samples from deviating from aviation facts, the system uses graph constraints to validate the samples, such as verifying whether the aircraft type matches the airport runway conditions, whether the procedure sequence conflicts, and whether the operational restrictions are still in effect; if a sample does not meet the constraints, it is added to the counterexample set. For samples that pass the validation, the system further establishes anchor mappings between questions or answers and manual paragraphs, drawing areas, table items, or audio clips, so as to explicitly learn the correspondence between answers and evidence during training.

[0037] In one embodiment, the joint training module uses a dual-channel structure. The question encoding channel receives token sequences of user questions or questions in the augmented training set and outputs a semantic representation of the question. The evidence encoding channel receives candidate subgraphs, source fragment vectors, image region vectors, and real-time data features corresponding to the question and outputs an evidence representation. The interactive fusion unit aligns the semantic representation of the question with the evidence representation, which can be achieved using a cross-attention mechanism or a gated fusion mechanism. Four types of losses are used jointly during training: first, answer generation loss, used to improve the correctness of the model's answers; second, evidence alignment loss, used to improve the matching degree between the answer and the cited evidence; third, time validity constraint loss, used to suppress the citation of invalid evidence; and fourth, safe rejection loss, used to train the model to output supplementary confirmation, risk warning, or rejection results when there is insufficient evidence or the question involves high-risk operations. Through this collaborative training, the model maintains consistency with structured evidence while expressing naturalness.

[0038] In one embodiment, the interpretable question-answering engine performs candidate evidence consistency scoring before generating an answer. The candidate evidence consistency score can be obtained by a linear or non-linear combination of the following scoring items: graph structure matching degree, source reliability, time validity, and real-time data consistency. Graph structure matching degree measures the degree of matching between the candidate graph path and the question intent; source reliability measures the level and historical stability of the evidence source; time validity measures whether the candidate evidence is within its validity period; and real-time data consistency measures whether there is a conflict between the real-time state and the static graph. If multiple candidate pieces of evidence conflict, expired evidence is first eliminated, and then evidence with higher source reliability is retained; if multiple candidate pieces of evidence still exist, evidence with stricter constraints is retained first. Only when the final candidate evidence consistency score is higher than a threshold does the system output a conclusive answer; otherwise, it outputs a supplementary confirmation prompt, a risk warning, or a rejection result.

[0039] In one embodiment, the dynamic knowledge update process is as follows: Figure 3 As shown, the system continuously monitors data sources such as operation notices, weather messages, ADS-B status, airport operation status, and maintenance work cards. When a new, modified, or failed event is detected, the event monitoring unit first performs structured parsing of the changed content and performs only a local incremental update in the shadow graph, without immediately overwriting the official graph. Subsequently, the system automatically generates simulated questions and answers around the changed entities, relationships, or constraints, such as new questions about newly released operation restrictions, weather restrictions, or maintenance procedures. The adapter incremental learning unit updates the lightweight parameters in the question-and-answer model based on the simulated questions and answers. During the verification phase, answer consistency, evidence citation integrity, and security rule compliance are calculated simultaneously. When the verification results reach a preset threshold, the shadow graph and the updated model are switched to the official version; if any indicator fails to meet the threshold, a rollback is performed, and the original official version continues to be used. This process can reduce the impact of real-time knowledge updates on system stability.

[0040] In one embodiment, an interpretable question-answering engine such as Figure 4As shown, the process includes problem classification and routing, problem decomposition, evidence retrieval, answer generation, consistency verification, and result presentation. The problem classification and routing unit first determines whether the user's problem falls under the categories of fact query, process handling, constraint judgment, comparative analysis, or dynamic situational problem. If it's a fact query, it prioritizes retrieving one-hop or multi-hop relationships from the graph and directly generating the answer based on the source fragment. If it's a process handling or constraint judgment, it first reasons along the process chain, constraint chain, and rule chain in the graph, and then the answer generation unit organizes natural language output. If it's a dynamic situational problem, it needs to combine both the formal graph and real-time data fragments to provide a response. The result presentation includes at least the answer text and one traceable piece of evidence, preferably also including the graph path, source fragment, time validity description, confidence level description, and highlighted areas in the diagram or table.

[0041] The following is a specific application example to illustrate the working process of this application. A user asks: "Given that crosswinds are increasing at the current destination and an alternate airport has just issued an operations notice (OBF), which alternate airport should be prioritized for a B737-800?" The system first identifies this question as a composite problem involving dynamic situational awareness and constraint judgment, and retrieves subgraphs related to aircraft type operational constraints, airport runway parameters, alternate landing rules, weather restrictions, and OBFs from the geographic data center. Simultaneously, it retrieves the latest meteorological information, operational status, and OBF summaries for candidate airports from real-time data sources. Subsequently, the system calculates the consistency score of each candidate piece of evidence and filters conflicting OBFs or weather information based on execution time validity and source reliability. If an OBF for a candidate airport has been updated but has not yet been published through shadow verification, the consistency score of the candidate evidence is below a threshold, and the system prompts "Relevant operational constraints are being verified; please manually review" in the answer instead of directly outputting a definitive recommendation. If the meteorological, runway parameter, and OBF for a candidate airport have all passed verification, the system outputs the preferred alternate airport and the reasons for the recommendation. The results not only provide recommended conclusions, but also the reasoning path of "aircraft type restrictions → airport parameters → weather conditions → operational notices", the cited notice excerpts, the time validity explanations of each candidate airport, and the confidence level explanations.

[0042] Let's take a maintenance assistance scenario as an example. A user asks, "After a low-pressure alarm occurs in the hydraulic system of a certain model, is it permissible to continue releasing the machine according to the existing configuration?" The system first searches for the association between the alarm code and the hydraulic system, minimum equipment list restrictions, fault isolation manual handling steps, and recent maintenance records. Then, it combines the constraints in the diagram to determine whether the conditions for continued release are met, and outputs the conclusion, the basis for the constraints, and the procedures requiring additional checks. If the retrieved constraints have version conflicts, effective time conflicts, or the evidence consistency score is below the threshold, the system will trigger a supplementary confirmation mechanism instead of directly giving a positive conclusion.

[0043] In one embodiment, the system can be implemented using a hybrid storage approach combining graph database, vector database, and relational database. The graph database stores the graph structure and multi-hop paths, the vector database stores question vectors, evidence vectors, and image region vectors, and the relational database stores source metadata, version information, and update logs. The lightweight adapter can be implemented using low-rank incremental parameters, thereby reducing the training overhead for each dynamic update. For the question classification, relation extraction, evidence alignment, and risk identification modules, this application does not limit the specific model type; any model that can implement the graph-model complementarity enhancement, dual-channel consistency training, dynamic shadow update mechanism, and evidence consistency control described in this application falls within the scope of protection of this application.

[0044] In summary, this application introduces a multimodal aviation knowledge graph with unified identifiers, time validity, and source fingerprints. Combined with graph constraint-driven large language model sample expansion, dual-channel evidence consistency training, shadow graph incremental update and rollback release mechanism, and interpretable question-answering output method, it can achieve more reliable, real-time, and traceable intelligent question-answering services in aviation professional scenarios.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent question-answering system for aviation based on graph schema synergistic enhancement, characterized in that, The system includes a data access and governance module, a multimodal aviation knowledge graph construction module, a graph-modal complementary data enhancement module, a joint training module, a dynamic knowledge update module, and an interpretable question-answering engine. The data access and governance module accesses aviation operation regulations, aircraft manufacturer technical data, maintenance manuals, fault records, operational notices, airport operation information, meteorological information, flight status information, voice recordings, drawings, images, and video data, and performs data cleaning, deduplication, standardization, unified identifier mapping, timestamp alignment, and anonymization processing on the data. The multimodal aviation knowledge graph construction module constructs an aviation knowledge graph containing entities, relationships, attributes, constraints, process chains, and valid time intervals from the processed data, and writes source fingerprints, evidence location information, and confidence levels for entities or relationships. The graph-modal complementary data enhancement module is used to... The aviation knowledge graph generates seed question-answer pairs and, combined with a large language model, generates rewritten samples, counterfactual samples, comparison samples, distractor samples, and safety constraint samples. After graph constraint verification, an enhanced training set is formed. The joint training module is used to perform dual-channel collaborative training on the question-answering model based on the enhanced training set and graph subgraph evidence to obtain an aviation domain question-answering model. The dynamic knowledge update module is used to perform incremental updates of the shadow graph, simulated question-answer generation, incremental learning of the adapter, and release rollback control when real-time data changes are detected. The interpretable question-answering engine is used to receive user questions, calculate the consistency score of candidate evidence, and output the answer text and its graph path, source fragment, time validity description, and confidence description when the consistency score meets the threshold. When the consistency score does not meet the threshold, it outputs supplementary confirmation prompts, risk warnings, or rejection results.

2. The graph-module-cooperative-augmented-aeronautical-intelligent-question-answering system of claim 1, wherein, The multimodal aviation knowledge graph construction module includes a text parsing unit, an image parsing unit, a voice parsing unit, an entity relationship extraction unit, and a timeliness annotation unit. The text parsing unit extracts aircraft type, system, component, fault mode, constraint, handling action, airport, runway, meteorological phenomenon, and operational event from flight manuals, maintenance manuals, minimum equipment lists, configuration deviation lists, fault handling records, and operational notices. The image parsing unit extracts component numbers, tag number mappings, and structural associations from design drawings, component diagrams, and alarm interface diagrams. The voice parsing unit extracts event context from maintenance, flight, or command communication text. The timeliness annotation unit records the effective time, expiration time, data version, source fingerprint, and confidence level for entities, relationships, or attributes to form an aviation knowledge graph capable of time-constrained reasoning.

3. The graph-module-cooperative-augmented-aeronautical-intelligent-question-answering system of claim 1, wherein, The graph-template complementary data enhancement module includes a graph template sampling unit, a semantic expansion unit, a constraint verification unit, and a counterexample generation unit. The graph template sampling unit generates seed question-answer pairs for fact query, process handling, constraint judgment, comparative analysis, and dynamic situation categories based on entity type, relationship path, process chain, and constraints. The semantic expansion unit generates term substitution samples, omission samples, cross-language samples, counterfactual samples, and distractor samples based on different questioning intentions. The constraint verification unit filters inconsistent samples based on graph relationship constraints, aviation safety rule base, and valid time intervals. The counterexample generation unit writes expanded samples that violate aircraft type restrictions, airport restrictions, weather restrictions, process sequence, or airworthiness rules into the counterexample set for subsequent rejection training.

4. The graph-module-cooperative-augmented-aeronautical-intelligent- question-answering system of claim 1, wherein, The joint training module includes a question encoding channel, an evidence encoding channel, an interactive fusion unit, and a consistency optimization unit. The question encoding channel is used to semantically encode user questions and questions in the augmented training set. The evidence encoding channel is used to structurally encode candidate graph subgraphs, source fragments, graph regions, table entries, and real-time data fragments. The interactive fusion unit achieves alignment between question representations and evidence representations through cross-attention or gating fusion mechanisms. The consistency optimization unit jointly optimizes the answer generation loss, evidence alignment loss, time validity constraint loss, and safe rejection loss, so that the answer text is consistent with the graph path and source evidence on which it depends.

5. The graph-module-cooperative-augmented-aeronautical-intelligent- question-answering system of claim 1, wherein, The interpretable question-answering engine includes a question classification and routing unit, a candidate evidence recall unit, an evidence consistency scoring unit, an answer generation unit, and an evidence presentation unit. The evidence consistency scoring unit calculates the candidate evidence consistency score based on graph structure matching degree, source reliability, time validity, and real-time data consistency. When multiple candidate pieces of evidence conflict with each other, they are filtered in the order of time validity priority, source reliability priority, and stricter constraints priority. When the consistency score of the filtered candidate evidence is lower than the threshold, the answer generation unit outputs a supplementary confirmation prompt, a risk warning, or a rejection result.

6. The image-model collaborative enhanced aviation intelligent question-answering system as described in claim 1, characterized in that, The dynamic knowledge update module includes an event monitoring unit, a shadow graph update unit, a simulated question-and-answer generation unit, an adapter incremental learning unit, and a release rollback unit. The event monitoring unit is used to monitor aviation operation notices, weather messages, ADS-B status, maintenance work cards, and new fault records. The shadow graph update unit is used to perform incremental updates in the shadow graph first without directly overwriting the official graph. The simulated question-and-answer generation unit is used to generate verification questions and answers around newly added or changed nodes, relationships, and constraints. The adapter incremental learning unit is used to perform lightweight parameter updates based on simulated question answering; The release rollback unit is used to release updates when the consistency index, accuracy index, and security verification index meet preset thresholds; otherwise, the original official map and model version are retained.

7. An intelligent question and answer method for aviation based on graph and model synergy, characterized in that, The system, applied to any one of claims 1 to 6, comprises the following steps: S1, collecting structured data, semi-structured data, unstructured data, and multimodal data in the aviation field, and cleaning, standardizing, mapping with unified identifiers, aligning with timestamps, and labeling with timeliness attributes on the data; S2, extracting entities, relationships, attributes, constraints, and process chains from the processed aviation data, and constructing a multimodal aviation knowledge graph with unified identifiers, valid time intervals, source fingerprints, evidence location information, and confidence levels; S3, generating seed question-answer pairs based on the entity relationship paths, process chains, and constraint rules of the multimodal aviation knowledge graph, and using a large language model to generate rewritten samples, counterfactual samples, comparison samples, interference samples, and safety constraint samples. S4. Then, an enhanced training set is obtained through graph rules, time constraints, and evidence consistency verification. S5. The question representation and the graph subgraph evidence representation are input into the joint training module, and joint optimization of answer generation loss, evidence alignment loss, time validity constraint loss, and safety rejection loss is performed to obtain the aviation domain question answering model. S6. After receiving the user's question, the question is first classified and routed, then the corresponding graph subgraph, source fragment, and real-time data are recalled, and the candidate evidence consistency score is calculated. S7. When the candidate evidence consistency score meets the threshold, the answer text, graph path, source fragment, time validity description, and confidence description are output. When the candidate evidence consistency score does not meet the threshold, a supplementary confirmation prompt, risk warning, or rejection result is output.

8. The graph-module synergistic enhanced aerial intelligence question answering method of claim 7, wherein, Step S3 further includes: generating safety constraint samples for issues involving operational restrictions, maintenance release, emergency response, and airworthiness constraints; when the samples obtained by expanding the large language model violate aircraft type restrictions, airport restrictions, weather restrictions, process sequence relationships, or airworthiness rules, the corresponding samples are marked as invalid samples and written into the rule counterexample set for use as rejection constraints in subsequent training.

9. The graph-module synergistic enhanced aerial intelligence question answering method of claim 7, wherein, When calculating the consistency score of candidate evidence in step S5, at least three of the following factors are considered: the matching degree between the candidate graph path and the semantics of the problem, the matching degree between the candidate source fragment and the anchor point of the graph path, the reliability level of the evidence source, the matching degree between the evidence effective time and the current time, and the degree of conflict between real-time data and static graph; and when there is conflicting evidence, priority is given to retaining candidate evidence that is still valid and has a higher reliability level.

10. The image-model collaborative enhancement aviation intelligent question answering method as described in claim 7, characterized in that, Before step S5, the method further includes: when changes are detected in operation notices, meteorological information, flight status information, maintenance work cards, or fault records, incremental updates are first performed in the shadow map and simulated questions and answers for the change points are generated, and then a lightweight adapter is used to incrementally learn the aviation domain question and answer model; when the simulated questions and answers pass the verification, the updated map and model are switched to the official version, otherwise a rollback is performed.

Images