Airport Operation Semantic Understanding and Command Conflict Resolution System Based on Large Language Model

Abstract

This invention discloses an airport operation semantic understanding and instruction conflict resolution system based on a large language model, belonging to the field of natural language processing technology. It includes: a data preprocessing module, a multi-level deep feature extraction module, a domain-specific large language model training and fine-tuning module, an airport operation instruction semantic understanding module, an instruction conflict resolution module, and an airport operation domain knowledge base. The data preprocessing module is used to aggregate, clean, normalize, label, and partition historical airport operation instruction data from multiple sources, outputting a standardized text dataset. This invention constructs a progressive vertical deep feature extraction architecture, realizing the step-by-step deep mining of airport operation instructions from basic semantic entities and syntactic features, single-instruction scenario-based business logic features, to cross-subject instruction interaction global dependencies and conflict-sensitive features, completely solving the shortcomings of the flat feature extraction method of general large language models that cannot adapt to the vertical domain of airport operations.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to an airport operation semantic understanding and instruction conflict resolution system based on a large language model. Background Technology

[0002] Airports are complex, multi-entity collaborative systems. Their daily operations involve air traffic control, airport operations command centers, airlines, ground handling companies, security, and emergency response, among others. Operational instructions take various forms, including voice-to-text, written dispatch instructions, system work orders, and walkie-talkie text. These instructions are characterized by strong domain specialization, numerous colloquial variations, high timeliness requirements, complex boundaries of responsibilities among stakeholders, and high coupling between instructions. If semantic misunderstandings or instruction conflicts are not resolved in a timely manner, they can lead to flight delays and reduced operational efficiency, or even serious safety hazards such as ground taxiing conflicts and apron accidents.

[0003] Existing technologies for addressing airport operational conflicts primarily focus on physical-level scheduling optimization, such as taxiway planning, parking space allocation, and flight path conflict detection. Examples include aircraft conflict detection based on road network models and joint scheduling using multi-agent reinforcement learning. These solutions rely on structured operational data and cannot handle unstructured natural language commands issued by multiple agents, nor can they resolve core issues such as semantic ambiguity and implicit conflicts inherent in the commands themselves.

[0004] In the application of semantic understanding technology, existing solutions mostly use general large language models for basic text parsing, which has the following core shortcomings: The general-purpose large language model lacks deep adaptation to the vertical field of airport operations, has low accuracy in understanding professional terms and scenario-based sentences, and is prone to entity recognition errors and semantic ambiguity parsing errors, thus failing to meet the high security requirements of airport operations. Most semantic feature extraction methods have a flat structure, which cannot realize the step-by-step in-depth mining of airport operation instructions from basic entities to business logic, and then to cross-subject global conflict features. The feature extraction granularity is insufficient, and it is impossible to accurately capture the implicit conflicts between instructions, resulting in a high rate of missed and false judgments in conflict detection. The command conflict resolution solutions mostly rely on manually preset rule bases, which have poor generalization ability and cannot cope with atypical conflicts in complex airport scenarios. In addition, the rule updates are lagging behind and cannot adapt to the high concurrency and low latency requirements of real-time airport operations. The inability to deeply integrate semantic understanding with conflict resolution means that semantic parsing results cannot directly support the accurate location and resolution of conflicts. The lack of an end-to-end integrated processing system makes engineering implementation difficult. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide an airport operation semantic understanding and instruction conflict resolution system based on a large language model; it can solve the problems that general large language models lack deep adaptation to the vertical field of airport operations, semantic feature extraction methods are mostly flat, instruction conflict resolution schemes mostly rely on manually preset rule bases, and cannot deeply integrate semantic understanding and conflict resolution.

[0006] Technical solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, an airport operation semantic understanding and instruction conflict resolution system based on a large language model, comprising: a data preprocessing module, a multi-level deep feature extraction module, a domain large language model training and fine-tuning module, an airport operation instruction semantic understanding module, an instruction conflict resolution module, and an airport operation domain knowledge base; Data preprocessing module: used to aggregate, clean, normalize, label and divide the airport operation historical instruction data from multiple sources, and output a standardized text dataset; Multi-level deep feature extraction module: used to perform a third-level progressive deep feature extraction on standardized text data. The second-level feature extraction takes the output of the first-level feature extraction as its only input, and the third-level feature extraction takes the output of the second-level feature extraction as its only input. The final output is a global deep feature vector used for model training and inference. Domain-specific large language model training and fine-tuning module: used to perform incremental pre-training and instruction fine-tuning of pre-trained large language models based on global deep feature vectors, and to build an end-to-end large model for airport operation semantic understanding and instruction conflict resolution; Airport Operation Instruction Semantic Understanding Module: This module is used to preprocess real-time input airport operation instructions, extract third-level features, and perform model reasoning, outputting structured semantic parsing results. Instruction conflict resolution module: Based on semantic parsing results and global deep features, it completes the detection, location, level determination and resolution scheme generation of instruction conflicts, realizing a closed loop of conflict handling; Airport Operations Knowledge Base: Provides fundamental support for the entire process by including civil aviation regulations, airport operation standards, professional terminology, business rules, and conflict type systems.

[0007] Furthermore, the specific processing flow of the data preprocessing module includes: Multi-source data aggregation: Aggregates historical airport operation data, including tower control voice transcription text, airport operation command center dispatch instruction text, ground support work order text, airline release instruction text, apron operation walkie-talkie call transcription text, emergency response instruction text, flight dynamic data, and airport resource status data; Data cleaning: Remove invalid data, including duplicate text, meaningless interjections, noisy characters, incomplete sentences, and filter text content that is irrelevant to airport operations; Text normalization: Standardize and unify the format of technical terms, flight numbers, gate numbers, runway numbers, taxiway numbers, time expressions, and spatial location expressions; Data annotation: Based on airport operation domain standards and business rules, text data is annotated with domain entities, scene tags, instruction types, and conflict tags; Dataset partitioning: The labeled dataset is divided into training set, validation set, and test set according to a preset ratio.

[0008] Furthermore, the multi-level deep feature extraction module includes a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit cascaded in sequence. The first-level feature extraction unit is used to extract basic semantic entities and syntactic features in the airport operation domain. The input is preprocessed standardized text, and the output is the first-level basic semantic feature vector matrix. The second-level feature extraction unit is used to perform airport operation scenario-based business logic feature extraction, and the output is the second-level scenario-based business logic feature vector matrix; The third-level feature extraction unit is used to perform cross-subject instruction interaction global dependency and conflict-sensitive feature extraction, and the output is a third-level global deep feature vector.

[0009] Furthermore, the basic semantic entity and syntactic feature extraction performed by the first-level feature extraction unit specifically includes the following steps: Step 1: Domain Terminology and Entity Embedding Encoding: Based on the airport operation domain knowledge base, an airport-specific terminology lexicon and entity dictionary are constructed. The entity types in the entity dictionary include subject, resource, action, spatiotemporal constraint, and state types. An extended embedding is performed using a pre-trained large language model's token embedding layer combined with the domain terminology lexicon. After tokenizing the input text, an initial token embedding vector is generated. The named entity recognition sub-model is used to identify and classify airport domain entities in the text. Entity type masks are added to the identified entities to generate entity-aware embedding vectors. The initial token embedding vector and the entity-aware embedding vector are concatenated to obtain entity-enhanced word-level embedding features. The second step, syntactic dependency and sentence structure feature encoding, involves using a syntactic dependency analysis model to perform syntactic parsing on the input text, extracting subject-verb-object structures, modifier relationships, core predicates, and syntactic relationships between the initiator and receiver of the instruction, and generating a syntactic dependency feature matrix. Based on a pre-built airport instruction sentence structure template library, the sentence structure types of the input text are matched and classified to generate sentence structure type encoding vectors. The syntactic dependency feature matrix and the sentence structure type encoding vectors are then fused to obtain syntactic-level feature vectors. The third step, first-level feature fusion and normalization, involves cross-dimensional fusion of entity-enhanced word-level embedding features and syntactic-level feature vectors. A multi-head attention mechanism is used to assign weights to the fused features, strengthening the feature weights of core entities and core syntactic structures while weakening the weights of meaningless modifiers. Through layer normalization and linear transformation, a fixed-dimensional first-level basic semantic feature vector matrix is ​​generated.

[0010] Furthermore, the scenario-based business logic feature extraction performed by the second-level feature extraction unit specifically includes the following steps: Step 1: Business Scenario Classification and Scenario Feature Anchoring: A scenario classification system is constructed based on the entire business process of airport operations. The scenario classification system includes eight major categories of core scenarios: takeoff control, landing control, ground taxiing, apron support, passenger service, emergency dispatch, flight release, and resource dispatch. Each major category is further subdivided into corresponding sub-scenarios. Based on the first-level basic semantic feature vector, a scenario classification sub-model is used to classify the business scenarios of input instructions, and the scenario classification results and scenario probability distribution vectors are output. Based on the pre-built business rule library for each scenario, the rule embedding vectors of the core business elements, constraint rules, and rights and responsibilities boundaries of the corresponding scenario are extracted. The scenario probability distribution vector and the rule embedding vector are concatenated to generate a scenario-aware rule feature vector. The second step, core business element extraction and association modeling of instructions: Based on the domain entities and syntactic structures identified by the first-level feature recognition, and combined with the rule feature vectors of scene awareness, the six core business elements of a single instruction are extracted: instruction initiating subject, instruction receiving subject, instruction execution action, execution object, spatiotemporal constraints, and execution conditions. A "subject-action-object-constraint" business element quadruple is constructed. A graph neural network is used, with each business element as a node and the business relationship between elements as an edge, to perform association modeling on the business element quadruple, generating business element association graph features. Through the business rule verification sub-model, the business compliance between elements is initially verified, and compliance feature codes are generated. The business element association graph features and compliance feature codes are fused to obtain the deep features of the business elements. The third step, instruction intent and type deep encoding: Based on the first-level sentence structure features and business element deep features, the core intent and instruction type of the instruction are classified and encoded to generate instruction intent encoding vector and instruction type encoding vector. These are then concatenated with the business element deep features to obtain instruction-level business semantic features. Step 4, Second-level feature fusion and dimensional alignment: The scene-aware rule feature vector, business element deep features, and instruction-level business semantic features are fused in multiple dimensions. A cross-attention mechanism is used to achieve deep interaction between the first-level basic semantic features and the second-level business logic features, strengthen the feature weights that conform to the current scene business rules, and filter out noisy features that are irrelevant to the business logic. Through layer normalization and linear transformation, a second-level scene-based business logic feature vector matrix that is aligned with the dimensions of the first-level feature vector is generated.

[0011] Furthermore, the cross-subject instruction interaction global dependency and conflict-sensitive feature extraction performed by the third-level feature extraction unit specifically includes the following steps: Step 1: Modeling the timing and subject grouping features of instructions: For all instructions within the same preset time window, sort them according to their timestamps to generate an instruction timing sequence; Based on the instruction initiating subject and receiving subject information in the second-level features, group the instructions into subjects to generate subject grouping features; Use a timing coding model to encode the timing position, time window, and subject affiliation of the instructions to generate a timing-subject-aware coding vector, which is concatenated with the input second-level scenario-based business logic feature vector matrix to obtain a timing-subject-enhanced multi-instruction feature matrix; Step 2: Cross-instruction global dependency modeling: For the multi-instruction feature matrix with temporal-subject enhancement, a multi-head cross-attention mechanism is used to construct a cross-instruction global dependency graph. Each instruction is a node, and the business relationships between instructions are edges. These business relationships include temporal dependencies, subject responsibility relationships, resource consumption relationships, execution condition dependencies, and action mutual exclusion relationships. A graph Transformer network is used to perform deep encoding on the global dependency graph to mine the implicit dependencies between instructions and generate a cross-instruction global dependency feature matrix. Step 3: Conflict-Sensitive Feature Anchoring and Enhancement: Based on a pre-built airport operation instruction conflict type library, conflict-sensitive sites are anchored in the global dependency feature matrix. The conflict type library includes spatiotemporal conflicts, action conflicts, responsibility conflicts, temporal conflicts, and conditional conflicts. A contrastive learning mechanism is adopted to compare the features of positive samples labeled with conflict and negative samples without conflict, learn the difference distribution between conflict features and non-conflict features, strengthen the feature weights of conflict-sensitive sites, and weaken features unrelated to conflict. Conflict level encoding vectors are generated for different levels of conflict and fused with the anchored conflict-sensitive features to obtain global features enhanced by conflict. Step 4, Third-level feature aggregation and global feature output: Deeply fuse the multi-instruction features enhanced by temporal-subject, the global dependency features across instructions, and the global features enhanced by conflict. Use gated recurrent units to aggregate the temporal sequence features. Combining global average pooling and max pooling, the fused features are compressed and aggregated in dimension. After layer normalization and linear transformation, a fixed-dimensional third-level global deep feature vector is generated.

[0012] Furthermore, the specific processing flow of the domain-specific large language model training and fine-tuning module includes: Base model selection and initialization: A general pre-trained large language model is selected as the base model, its pre-trained weights are loaded, and the embedding layer, attention layer and output layer of the model are adapted to make it suitable for the feature vector dimension output by the multi-level deep feature extraction module. Incremental pre-training stage: Based on the full corpus data of airport operations, combined with the first, second and third level features output by the multi-level deep feature extraction module, the base model is incrementally pre-trained to enable the model to learn the professional knowledge, terminology system, business rules and semantic logic of airport operations. Instruction fine-tuning stage: Based on the labeled instruction fine-tuning dataset, the global deep feature vector output by the multi-level deep feature extraction module is used as the core input, and the instruction semantic parsing results, conflict detection results, conflict level determination, and conflict resolution schemes are used as the output targets. The model is fine-tuned in a supervised manner, and the conflict resolution scheme generation ability of the model is optimized by combining human feedback reinforcement learning. Model Validation and Optimization: The trained model is validated based on the test set. Validation metrics include semantic understanding accuracy, entity recognition accuracy, scene classification accuracy, conflict detection precision, recall, F1 score, and compliance rate of conflict resolution solutions. The weights of the multi-level deep feature extraction module, model parameters, and dataset are optimized and iterated to address validation deviations until the model performance meets the safety and efficiency requirements of airport operations.

[0013] Furthermore, the specific processing flow of the airport operation instruction semantic understanding module includes: Real-time instruction preprocessing: Perform preprocessing operations consistent with historical data on real-time input speech-to-text, dispatch instruction text, work order text, and walkie-talkie conversation text to generate standardized text; Real-time extraction of third-level features: Through the multi-level deep feature extraction module, the standardized text is extracted step by step to generate the corresponding first-level basic semantic features, second-level scenario-based business logic features, and third-level global deep features; Model reasoning and semantic parsing: Input the third-level global deep features into the trained domain language model, and the model outputs standardized semantic parsing results, including the instruction core element quadruple, business scenario, instruction type, instruction intent, compliance verification results, and entity standardization results; Semantic result output: The semantic parsing results are stored in a structured manner and pushed synchronously to the instruction conflict resolution module.

[0014] Furthermore, the specific processing flow of the instruction conflict resolution module includes: Conflict detection and localization: Based on the conflict-sensitive features in the third-level global deep features, combined with the semantic parsing results of all instructions within the current time window, instruction conflicts are detected through model inference; if a conflict exists, the type of conflict, the instructions involved, the conflict subject, the conflict location, and the scope of impact are accurately located. Conflict Level and Risk Assessment: Based on a pre-built airport operation safety risk level system, and combined with the conflict type, scope of impact, and urgency, the conflict is classified into four levels: low risk, medium risk, high risk, and extremely high risk. The potential safety risks and operational impacts of the conflict are assessed simultaneously. Conflict resolution solution generation: For conflicts of different levels, multiple compliant conflict resolution solutions are generated by combining civil aviation regulations, airport operation standards, business rules, and current airport resource status through a domain-wide big language model. Each solution includes handling steps, responsible parties, execution time windows, resource allocation suggestions, and risk prevention and control measures. Solution optimization and push: Based on preset decision indicators, multiple resolution solutions are ranked and optimized. The decision indicators are ranked from high to low priority as follows: highest safety priority, least operational impact, highest execution efficiency, and least resource consumption. The optimal solution is output. The conflict information, risk level, and optimal resolution solution are simultaneously pushed to the corresponding responsible parties and the airport operation command center. Closed-loop management and model iteration: Track the implementation of conflict resolution solutions, record the results and feedback information, supplement the training dataset with the case data of completed cases, and perform incremental fine-tuning of the model on a regular basis.

[0015] Furthermore, the airport operations knowledge base includes a civil aviation regulations database, an airport operation specifications database, a professional terminology dictionary, a business rules database, a scenario classification system, a conflict type database, an emergency plan database, and a historical handling case database, supporting real-time updates and iterations. The system also includes a visualization and emergency alarm module, used to visualize semantic parsing results, conflict detection results, and resolution solutions, triggering corresponding alarms for different levels of conflict, and supporting full-process instruction traceability and model effect monitoring.

[0016] Beneficial effects: This invention constructs a progressive vertical deep feature extraction architecture, realizing the step-by-step deep mining of airport operation instructions from basic semantic entities and syntactic features, single-instruction scenario-based business logic features, to cross-subject instruction interaction global dependencies and conflict-sensitive features. It completely solves the defect that the flat feature extraction method of general large language models cannot be adapted to the vertical domain of airport operations, greatly improves the accuracy of domain semantic understanding, and accurately captures implicit conflicts between instructions, significantly reducing the false negative rate and false positive rate of conflict detection.

[0017] By deeply integrating multi-level deep feature extraction with large language models, an end-to-end airport operation semantic understanding and instruction conflict resolution system has been constructed. This system automates the entire process from unstructured instruction text input, standardized semantic parsing, conflict detection and localization, risk level determination to the generation of compliant resolution solutions. It does not rely on manually preset rule bases, has strong generalization capabilities, and can effectively deal with various typical and atypical instruction conflicts in complex airport operation scenarios. This significantly reduces the workload of manual verification and the risk of human error, and significantly improves the overall efficiency of daily airport operations.

[0018] In response to the professional, high-security, and multi-entity collaborative nature of airport operations, the entire process of feature extraction and model training has been adapted to the vertical domain. From terminology embedding and scenario rule embedding to conflict-sensitive feature enhancement, the entire process is aligned with the business logic and safety regulations of airport operations. The semantic parsing results and resolution solutions output by the model fully comply with civil aviation regulations and airport operation requirements, effectively avoiding safety hazards such as ground taxiing conflicts and apron safety accidents, and ensuring the safe and stable operation of the airport.

[0019] A closed-loop conflict resolution and model iteration optimization mechanism was constructed. By tracking the execution of conflict resolution solutions, recording the resolution results and on-site feedback information, and supplementing the model training dataset with complete case data of the resolved conflicts, the mechanism continuously optimizes the model performance by periodically performing incremental fine-tuning on the domain-specific large language model. At the same time, a real-time updatable airport operation domain knowledge base is provided to quickly adapt to the dynamic updates of airport operation business rules, scenario requirements and civil aviation regulations, and meet the high concurrency and low latency processing requirements of real-time airport operations. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system principle. Detailed Implementation

[0021] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Example 1

[0022] This embodiment is based on the airport operation semantic understanding and command conflict resolution system based on a large language model described in this invention. It addresses the ground taxiing and apron support operation scenarios in the daily operation of civil transport airports, completing the entire process of command semantic understanding and conflict resolution. The specific implementation process is as follows: System basic deployment and airport operations knowledge base initialization First, the hardware and software environment of the system is deployed, and a server cluster is deployed to meet the computing power requirements for model training and real-time inference. At the same time, the initial construction of the knowledge base in the field of airport operations is completed. The airport operations knowledge base includes a civil aviation regulations database, an airport operation specifications database, a professional terminology dictionary, a business rules database, a scenario classification system, a conflict type database, an emergency plan database, and a historical case database. The civil aviation regulations database contains currently effective civil aviation regulations and normative documents related to airport operation management. The airport operation specifications database contains the airport's ground taxiing management rules, apron operation safety management rules, and rules for the use of aircraft stands and taxiways. The professional terminology dictionary contains all professional terms and standardized expressions related to airport ground operations and apron support. The business rules database contains business processes, boundaries of responsibilities, and operational constraints in ground taxiing and apron support scenarios. The scenario classification system categorizes ground taxiing and apron support scenarios into major categories and corresponding sub-scenarios. The conflict type database contains the criteria for judging five major types of conflicts: spatiotemporal conflicts, action conflicts, responsibility conflicts, temporal conflicts, and conditional conflicts. The emergency plan database contains emergency response procedures for apron operation conflicts and ground taxiing conflicts. The historical case database contains case studies of command conflict handling related to ground taxiing and apron support at the airport over the past three years. The knowledge base supports real-time updates and iterations.

[0023] The data preprocessing module performs historical data processing to generate a standardized text dataset. The data preprocessing module performs full-process processing on the airport's historical operational instruction data for the past three years. The specific process is as follows: The first step is to aggregate multi-source data: aggregate tower control voice transcription text, airport operations command center dispatch instruction text, ground support work order text, airline release instruction text, apron operation walkie-talkie call transcription text, emergency response instruction text, flight dynamic data, and airport resource status data to complete the unified aggregation of multi-source heterogeneous data; The second step is data cleaning: removing duplicate text, meaningless interjections, noisy characters, and incomplete sentences from the aggregated data, filtering text content that is irrelevant to airport operations, and retaining valid operational instruction data. The third step is text normalization: standardize and unify the professional terms, flight numbers, gate numbers, runway numbers, taxiway numbers, time expressions, and spatial location expressions in the text. For example, the colloquial "gate number 12" is standardized as "gate 12", "3 pm" is standardized as "15:00" in 24-hour format, and the same professional term expressed differently is mapped to the standardized term in the dictionary. The fourth step is data annotation: Based on airport operation domain standards and business rules, the cleaned and normalized text data is annotated with domain entities, scene tags, instruction types, and conflict tags. Among them, entity annotation includes five types: subject, resource, action, spatiotemporal constraint, and state. Scene tag annotation corresponds to ground taxiing and apron support related scenarios. Instruction type annotation corresponds to control instructions, dispatch instructions, support work orders, etc. Conflict tags indicate whether the instructions conflict, the type of conflict, and the level of conflict. Step 5: Dataset partitioning: Divide the labeled dataset into training set, validation set, and test set according to a preset ratio of 8:1:1, and output the standardized text dataset to the multi-level deep feature extraction module.

[0024] The multi-level deep feature extraction module performs the third-level progressive deep feature extraction. The multi-level deep feature extraction module comprises a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit cascaded sequentially. It performs step-by-step feature extraction on a standardized text dataset. The second-level feature extraction uses the output of the first-level feature extraction as its sole input, and the third-level feature extraction uses the output of the second-level feature extraction as its sole input. The specific process is as follows: The first step involves the first-level feature extraction unit extracting basic semantic entities and syntactic features from the airport operations domain. The input is preprocessed standardized text, and the output is the first-level basic semantic feature vector matrix. The specific execution steps are as follows: 1. Domain Terminology and Entity Embedding Encoding: Based on the airport operation domain knowledge base, an airport-specific terminology lexicon and entity dictionary are constructed. The token embedding layer of a pre-trained large language model is combined with the domain terminology lexicon for extended embedding. After tokenizing the input text, an initial token embedding vector is generated. The named entity recognition sub-model is used to identify and classify airport domain entities in the text. Entity type masks are added to the identified entities to generate entity-aware embedding vectors. The initial token embedding vector and the entity-aware embedding vector are concatenated to obtain entity-enhanced word-level embedding features. 2. Syntactic Dependency and Sentence Pattern Feature Encoding: A syntactic dependency analysis model is used to parse the input text syntactically, extracting subject-verb-object structure, modifier relationships, core predicates, and syntactic relationships between the initiator and receiver of the instruction, generating a syntactic dependency feature matrix. Based on a pre-built airport instruction sentence pattern template library, the sentence pattern type of the input text is matched and classified, generating a sentence pattern type encoding vector. The syntactic dependency feature matrix and the sentence pattern type encoding vector are fused to obtain a syntactic-level feature vector. 3. First-level feature fusion and normalization: The word-level embedding features enhanced by entities are fused with syntactic feature vectors across dimensions. A multi-head attention mechanism is used to assign weights to the fused features, strengthening the feature weights of core entities and core syntactic structures and weakening the weights of meaningless modifiers. Through layer normalization and linear transformation, a fixed-dimensional first-level basic semantic feature vector matrix is ​​generated.

[0025] The second step involves the second-level feature extraction unit extracting airport operation scenario-based business logic features. The input is the first-level basic semantic feature vector matrix, and the output is the second-level scenario-based business logic feature vector matrix. The specific execution steps are as follows: 1. Business Scenario Classification and Scenario Feature Anchoring: Based on the scenario classification system built on the entire business process of airport operation, a scenario classification sub-model is used to classify the business scenarios of input instructions, and output the scenario classification results and scenario probability distribution vectors; based on the pre-built business rule library for each scenario, the rule embedding vectors of the core business elements, constraint rules, and rights and responsibilities boundaries of the corresponding scenario are extracted, and the scenario probability distribution vector and the rule embedding vector are concatenated to generate scenario-aware rule feature vectors; 2. Extraction and Association Modeling of Core Business Elements of Instructions: Based on the domain entities and syntactic structures identified by first-level feature recognition, and combined with the rule feature vectors of scene awareness, the six core business elements of a single instruction are extracted: instruction initiating subject, instruction receiving subject, instruction execution action, execution object, spatiotemporal constraints, and execution conditions. A "subject-action-object-constraint" business element quadruple is constructed. A graph neural network is used, with each business element as a node and the business relationship between elements as an edge, to perform association modeling on the business element quadruple, generating business element association graph features. Through the business rule verification sub-model, the business compliance between elements is initially verified, and compliance feature codes are generated. The business element association graph features and compliance feature codes are fused to obtain the deep features of the business elements. 3. Instruction Intent and Type Deep Encoding: Based on the first-level sentence structure features and business element deep features, the core intent and instruction type of the instruction are classified and encoded to generate instruction intent encoding vectors and instruction type encoding vectors. These are then concatenated with the business element deep features to obtain instruction-level business semantic features. 4. Second-level feature fusion and dimensional alignment: The scene-aware rule feature vector, business element deep features, and instruction-level business semantic features are fused in multiple dimensions. A cross-attention mechanism is used to achieve deep interaction between the first-level basic semantic features and the second-level business logic features, strengthen the feature weights that conform to the current scene business rules, and filter out noise features that are irrelevant to the business logic. Through layer normalization and linear transformation, a second-level scene-based business logic feature vector matrix that is aligned with the dimensions of the first-level feature vector is generated.

[0026] The third step involves the third-level feature extraction unit performing cross-subject instruction interaction global dependency and conflict-sensitive feature extraction. The input is the second-level scenario-based business logic feature vector matrix, and the output is the third-level global deep feature vector. The specific execution steps are as follows: 1. Command Timing and Subject Grouping Feature Modeling: For all commands within the same preset 5-minute time window, sort them according to their timestamps to generate a command timing sequence; based on the command initiating subject and receiving subject information in the second-level features, group the commands into subjects to generate subject grouping features; use a timing coding model to encode the timing position, time window, and subject attribution of the commands to generate a timing-subject-aware coding vector, which is concatenated with the input second-level scenario-based business logic feature vector matrix to obtain a timing-subject-enhanced multi-command feature matrix; 2. Cross-instruction global dependency modeling: For the multi-instruction feature matrix with temporal-subject enhancement, a multi-head cross-attention mechanism is used to construct a cross-instruction global dependency graph. Each instruction is a node, and the business relationships between instructions are edges. These business relationships include temporal dependencies, subject responsibility relationships, resource consumption relationships, execution condition dependencies, and action mutual exclusion relationships. A graph Transformer network is used to perform deep encoding on the global dependency graph to mine the implicit dependencies between instructions and generate a cross-instruction global dependency feature matrix. 3. Conflict-Sensitive Feature Anchoring and Enhancement: Based on a pre-built airport operation instruction conflict type library, conflict-sensitive sites are anchored on the global dependent feature matrix. A contrastive learning mechanism is adopted to compare the features of positive samples labeled with conflict and negative samples without conflict, learn the difference distribution between conflict features and non-conflict features, strengthen the feature weights of conflict-sensitive sites, and weaken features unrelated to conflict. Conflict level encoding vectors are generated for different levels of conflict and fused with the anchored conflict-sensitive features to obtain global features with enhanced conflict. 4. Third-level feature aggregation and global feature output: The temporal-subject-enhanced multi-instruction features, cross-instruction global dependency features, and conflict-enhanced global features are deeply fused. The temporal sequence features are aggregated using a gated recurrent unit. The fused features are compressed and aggregated by combining global average pooling and max pooling. After layer normalization and linear transformation, a fixed-dimensional third-level global deep feature vector is generated and output to the domain large language model training and fine-tuning module.

[0027] Domain-Specific Large Language Model Training and Fine-Tuning Modules for Building Airport Operation-Specific Large Models The domain-specific large language model training and fine-tuning module, based on global deep feature vectors, performs incremental pre-training and instruction fine-tuning on the pre-trained large language model, constructing an end-to-end large model for airport operation semantic understanding and instruction conflict resolution. The specific process is as follows: The first step is to select and initialize the base model: select an open-source general-purpose pre-trained large language model as the base model, load its pre-trained weights, and adapt the model's embedding layer, attention layer, and output layer to make it compatible with the feature vector dimension output by the multi-level deep feature extraction module. The second step, incremental pre-training stage: Based on the full corpus data of airport operations, combined with the first, second and third level features output by the multi-level deep feature extraction module, the base model is incrementally pre-trained to enable the model to learn the professional knowledge, terminology system, business rules and semantic logic of airport operations. The third step, instruction fine-tuning stage: Based on the labeled instruction fine-tuning dataset, the global deep feature vector output by the multi-level deep feature extraction module is used as the core input, and the instruction semantic parsing results, conflict detection results, conflict level determination, and conflict resolution schemes are used as the output targets. The model is fine-tuned in a supervised manner, and the model's conflict resolution scheme generation ability is optimized by combining human feedback reinforcement learning. The fourth step is model validation and optimization: The trained model is validated based on the test set. Validation metrics include semantic understanding accuracy, entity recognition accuracy, scene classification accuracy, conflict detection precision, recall, F1 score, and compliance rate of conflict resolution solutions. The weights of the multi-level deep feature extraction module, model parameters, and dataset are optimized and iterated to address validation deviations until the model performance meets the safety and efficiency requirements of airport operations, thus completing the construction of the dedicated large model.

[0028] The airport operation instruction semantic understanding module performs real-time instruction semantic parsing. For the two real-time input operational instructions, the first is a dispatch instruction text issued by the Airport Operations Command Center to the aircraft crew: "Flight CA1234, taxi out of stand 12, proceed via taxiway B and taxiway C to runway 05, arrive at the runway holding point before 15:00"; the second is a support work order text issued by the Airport Operations Command Center to the ground support department: "Ground Support Department Towing Team, from 14:55 to 15:05, perform faulty equipment towing operations on taxiway C section", the airport operational instruction semantic understanding module executes the following processing flow: The first step is real-time instruction preprocessing: the two real-time input scheduling instruction texts and support work order texts mentioned above are cleaned and normalized in accordance with historical data to generate standardized texts. The second step, real-time extraction of third-level features: Through the trained multi-level deep feature extraction module, the standardized text is extracted step by step to generate the corresponding first-level basic semantic features, second-level scenario-based business logic features, and third-level global deep features. The third step is model reasoning and semantic parsing: The third-level global deep features are input into the trained airport operation-specific big language model. The model outputs standardized semantic parsing results, including the core element quadruple of the two instructions, business scenario, instruction type, instruction intent, compliance verification results, and entity standardization results. The business scenario of the first instruction is ground taxiing, and the business scenario of the second instruction is apron support. The entity standardization results clearly show that both instructions involve the core resource of taxiway C, and the time windows overlap. The fourth step is semantic result output: the semantic parsing results are stored in a structured manner and pushed synchronously to the instruction conflict resolution module.

[0029] The instruction conflict resolution module performs full-process conflict handling and closed-loop management. The instruction conflict resolution module, based on semantic parsing results and global deep features, completes the detection, location, severity determination, and resolution scheme generation of instruction conflicts, realizing a closed loop for conflict handling. The specific process is as follows: The first step, conflict detection and localization: Based on the conflict-sensitive features in the third-level global deep features, combined with the semantic parsing results of the two instructions within the current 5-minute time window, the instruction conflict was detected through model inference; the conflict type was accurately located as a spatiotemporal conflict and a resource occupation conflict, involving the two ground taxiing scheduling instructions and the apron support work order mentioned above, the conflicting parties being the flight crew of flight CA1234 and the ground support department's towing vehicle team, the conflict location being taxiway section C, the conflict time window being from 14:55 to 15:00, and the impact range being the ground taxiing order in the runway 05 direction and the normal takeoff of flight CA1234; The second step is to assess the conflict level and risk: Based on the pre-built airport operation safety risk level system, and combined with the conflict type, scope of impact and urgency, the conflict level is determined to be high-risk. The potential safety risks of the conflict are assessed as the risk of collision between aircraft and support vehicles caused by the occupation of the ground taxiway, and the operational impacts are the delay of flight CA1234 and the disruption of taxiway traffic order. The third step is to generate conflict resolution solutions: In response to this high-risk conflict, three compliant conflict resolution solutions are generated by combining civil aviation regulations, airport operation standards, business rules, and the current status of airport resources through a domain-specific big language model. Each solution includes handling steps, responsible parties, execution time windows, resource allocation suggestions, and risk prevention and control measures. The fourth step is the optimization and promotion of the solution: Based on the preset decision indicators, the three resolution solutions are ranked and optimized. The decision indicators are ranked from highest to lowest as follows: highest safety priority, least operational impact, highest execution efficiency, and least resource consumption. The optimal solution is as follows: adjust the working time window of the ground support department's towing vehicle team to 15:10 to 15:20, adjust the working route to detour through taxiway section D to the working point, and have flight CA1234 taxi according to the original instructions. At the same time, the airport operations command center's on-site supervision personnel will monitor the traffic order of taxiway section C throughout the process. The conflict information, risk level, and optimal resolution solution are simultaneously promoted to the airport operations command center, the flight CA1234 crew, and the ground support department's towing vehicle team, and at the same time, trigger the high-risk level alarm of the visualization and emergency alarm module. The fifth step is to close the loop and iterate the model: track the implementation of the conflict resolution plan, record the results and feedback from each party, and after confirming that the conflict has been completely resolved and there are no safety hazards or operational impacts, add the case data of the completed resolution to the training dataset and regularly perform incremental fine-tuning on the domain-specific large language model.

[0030] Visualization and emergency alarm modules are used in conjunction with each other. The visualization and emergency alarm module provides a full-process visualization of the semantic parsing results, conflict detection results, and resolution plan execution progress of this command. It triggers corresponding level audible and visual alarms for high-risk conflicts and supports the tracing of the entire command processing process and real-time monitoring of model performance. Example 2

[0031] This embodiment is based on the airport operation semantic understanding and command conflict resolution system based on a large language model described in this invention. It addresses the multi-entity collaborative scenario of takeoff control and flight release in the daily operation of civil transport airports, completing the entire process of command semantic understanding and conflict resolution. The specific implementation process is as follows: System basic deployment and airport operations knowledge base initialization First, the hardware and software environment of the system is deployed, and a server cluster that meets the requirements of high-concurrency real-time inference is deployed. At the same time, the initial construction and real-time updating of the knowledge base in the field of airport operations are completed. The airport operations knowledge base includes a civil aviation regulations database, an airport operation specifications database, a professional terminology dictionary, a business rules database, a scenario classification system, a conflict type database, an emergency plan database, and a historical case database. The civil aviation regulations database contains currently effective civil aviation air traffic management rules and flight release management-related civil aviation regulations. The airport operation specifications database contains the airport's takeoff control procedures, flight release management specifications, and runway and takeoff standard management rules. The professional terminology dictionary contains all professional terms and standardized expressions related to takeoff control and flight release. The business rules database contains business processes, multi-entity responsibility boundaries, and takeoff constraint rules under takeoff control, flight release, and flight scheduling scenarios. The scenario classification system categorizes takeoff control, flight release, and emergency scheduling scenarios and their corresponding sub-scenarios. The conflict type database contains the judgment criteria and classification rules for five major conflict categories: spatiotemporal conflict, action conflict, responsibility conflict, sequence conflict, and conditional conflict. The emergency plan database contains emergency response procedures for takeoff control conflicts and flight release anomalies. The historical case database contains case studies of instruction conflict handling related to takeoff control and flight release from the past three years.

[0032] The data preprocessing module performs historical data processing to generate a standardized text dataset. The data preprocessing module performs full-process processing on historical operational instructions related to takeoff control and flight release for the airport over the past three years. The specific process is as follows: The first step is to aggregate multi-source data: aggregate tower control voice transcription text, airport operations command center flight scheduling instruction text, airline operations control center release instruction text, apron operation walkie-talkie call transcription text, flight dynamic data, airport meteorological and runway status data, and airport resource status data to complete the unified aggregation of multi-source heterogeneous data; The second step is data cleaning: removing duplicate text, meaningless interjections, noisy characters, incomplete sentences, and radio interference from the aggregated data; filtering out text content that is irrelevant to airport takeoff control and flight release operations; and retaining valid operational instruction data. The third step is text normalization: standardize and unify the professional terms, flight numbers, runway numbers, takeoff standard codes, time expressions, meteorological condition expressions, and spatial location expressions in the text. For example, the colloquial "runway 36" is standardized as "runway 36", "ten minutes later" is standardized as a time expression based on the instruction timestamp, and the same professional terms and meteorological codes in different expressions are mapped to standardized expressions in the dictionary. The fourth step is data annotation: Based on airport operation domain standards and business rules, the cleaned and normalized text data is annotated with domain entities, scene tags, instruction types, and conflict tags. Among them, entity annotation includes five types: subject, resource, action, spatiotemporal constraint, and status. Scene tag annotation corresponds to takeoff control and flight release related scenarios. Instruction type annotation corresponds to control instructions, release instructions, scheduling instructions, etc. Conflict tags indicate whether the instructions conflict, the type of conflict, the location of conflict, and the level of conflict. Step 5: Dataset partitioning: Divide the labeled dataset into training set, validation set, and test set according to a preset ratio of 8:1:1, and output the standardized text dataset to the multi-level deep feature extraction module.

[0033] The multi-level deep feature extraction module performs the third-level progressive deep feature extraction. The multi-level deep feature extraction module comprises a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit cascaded sequentially. It performs step-by-step feature extraction on a standardized text dataset. The second-level feature extraction uses the output of the first-level feature extraction as its sole input, and the third-level feature extraction uses the output of the second-level feature extraction as its sole input. The specific process is as follows: The first step involves the first-level feature extraction unit extracting basic semantic entities and syntactic features from the airport operations domain. The input is preprocessed standardized text, and the output is the first-level basic semantic feature vector matrix. The specific execution steps are as follows: 1. Domain Terminology and Entity Embedding Encoding: Based on the airport operation domain knowledge base, an airport-specific terminology lexicon and entity dictionary are constructed. The token embedding layer of a pre-trained large language model is combined with the domain terminology lexicon for extended embedding. After tokenizing the input text, an initial token embedding vector is generated. The named entity recognition sub-model is used to identify and classify airport domain entities in the text. Entity type masks are added to the identified entities to generate entity-aware embedding vectors. The initial token embedding vector and the entity-aware embedding vector are concatenated to obtain entity-enhanced word-level embedding features. 2. Syntactic Dependency and Sentence Pattern Feature Encoding: A syntactic dependency analysis model is used to parse the input text syntactically, extracting subject-verb-object structure, modifier relationships, core predicates, and syntactic relationships between the initiator and receiver of the instruction, generating a syntactic dependency feature matrix. Based on a pre-built airport instruction sentence pattern template library, the sentence pattern type of the input text is matched and classified, generating a sentence pattern type encoding vector. The syntactic dependency feature matrix and the sentence pattern type encoding vector are fused to obtain a syntactic-level feature vector. 3. First-level feature fusion and normalization: The word-level embedding features enhanced by entities are fused with syntactic feature vectors across dimensions. A multi-head attention mechanism is used to assign weights to the fused features, strengthening the feature weights of core entities and core syntactic structures, and weakening the weights of meaningless modifiers and modifiers. Through layer normalization and linear transformation, a fixed-dimensional first-level basic semantic feature vector matrix is ​​generated.

[0034] The second step involves the second-level feature extraction unit extracting airport operation scenario-based business logic features. The input is the first-level basic semantic feature vector matrix, and the output is the second-level scenario-based business logic feature vector matrix. The specific execution steps are as follows: 1. Business Scenario Classification and Scenario Feature Anchoring: Based on the scenario classification system built on the entire business process of airport operation, a scenario classification sub-model is used to classify the business scenarios of input instructions, and output the scenario classification results and scenario probability distribution vectors; based on the pre-built business rule library for each scenario, the rule embedding vectors of the core business elements, constraint rules, and rights and responsibilities boundaries of the corresponding scenario are extracted, and the scenario probability distribution vector and the rule embedding vector are concatenated to generate scenario-aware rule feature vectors; 2. Extraction and Association Modeling of Core Business Elements of Instructions: Based on the domain entities and syntactic structures identified by first-level feature recognition, and combined with the rule feature vectors of scene awareness, the six core business elements of a single instruction are extracted: instruction initiating subject, instruction receiving subject, instruction execution action, execution object, spatiotemporal constraints, and execution conditions. A "subject-action-object-constraint" business element quadruple is constructed. A graph neural network is used, with each business element as a node and the business relationship between elements as an edge, to perform association modeling on the business element quadruple, generating business element association graph features. Through the business rule verification sub-model, the business compliance between elements is initially verified, and compliance feature codes are generated. The business element association graph features and compliance feature codes are fused to obtain the deep features of the business elements. 3. Instruction Intent and Type Deep Encoding: Based on the first-level sentence structure features and business element deep features, the core intent and instruction type of the instruction are classified and encoded to generate instruction intent encoding vectors and instruction type encoding vectors. These are then concatenated with the business element deep features to obtain instruction-level business semantic features. 4. Second-level feature fusion and dimensional alignment: The scene-aware rule feature vector, business element deep features, and instruction-level business semantic features are fused in multiple dimensions. A cross-attention mechanism is used to achieve deep interaction between the first-level basic semantic features and the second-level business logic features, strengthen the feature weights that conform to the current scene business rules, and filter out noise features that are irrelevant to the business logic. Through layer normalization and linear transformation, a second-level scene-based business logic feature vector matrix that is aligned with the dimensions of the first-level feature vector is generated.

[0035] The third step involves the third-level feature extraction unit performing cross-subject instruction interaction global dependency and conflict-sensitive feature extraction. The input is the second-level scenario-based business logic feature vector matrix, and the output is the third-level global deep feature vector. The specific execution steps are as follows: 1. Command Timing and Subject Grouping Feature Modeling: For all commands within the same preset 10-minute time window, sort them according to their timestamps to generate a command timing sequence; based on the command initiating subject and receiving subject information in the second-level features, group the commands into subjects to generate subject grouping features; use a timing coding model to encode the timing position, time window, and subject attribution of the commands to generate a timing-subject-aware coding vector, which is concatenated with the input second-level scenario-based business logic feature vector matrix to obtain a timing-subject-enhanced multi-command feature matrix; 2. Cross-instruction global dependency modeling: For the multi-instruction feature matrix with temporal-subject enhancement, a multi-head cross-attention mechanism is used to construct a cross-instruction global dependency graph. Each instruction is a node, and the business relationships between instructions are edges. These business relationships include temporal dependencies, subject responsibility relationships, resource consumption relationships, execution condition dependencies, and action mutual exclusion relationships. A graph Transformer network is used to perform deep encoding on the global dependency graph to mine the implicit dependencies between instructions and generate a cross-instruction global dependency feature matrix. 3. Conflict-Sensitive Feature Anchoring and Enhancement: Based on a pre-built airport operation instruction conflict type library, conflict-sensitive sites are anchored on the global dependent feature matrix. A contrastive learning mechanism is adopted to compare the features of positive samples labeled with conflict and negative samples without conflict, learn the difference distribution between conflict features and non-conflict features, strengthen the feature weights of conflict-sensitive sites, and weaken features unrelated to conflict. Conflict level encoding vectors are generated for different levels of conflict and fused with the anchored conflict-sensitive features to obtain global features with enhanced conflict. 4. Third-level feature aggregation and global feature output: The temporal-subject-enhanced multi-instruction features, cross-instruction global dependency features, and conflict-enhanced global features are deeply fused. The temporal sequence features are aggregated using a gated recurrent unit. The fused features are compressed and aggregated by combining global average pooling and max pooling. After layer normalization and linear transformation, a fixed-dimensional third-level global deep feature vector is generated and output to the domain large language model training and fine-tuning module.

[0036] Domain-Specific Large Language Model Training and Fine-Tuning Modules for Building Airport Operation-Specific Large Models The domain-specific large language model training and fine-tuning module, based on global deep feature vectors, performs incremental pre-training and instruction fine-tuning on the pre-trained large language model, constructing an end-to-end large model for airport operation semantic understanding and instruction conflict resolution. The specific process is as follows: The first step is to select and initialize the base model: a general pre-trained large language model is selected as the base model, its pre-trained weights are loaded, and the embedding layer, attention layer and output layer of the model are adapted to make it suitable for the feature vector dimension output by the multi-level deep feature extraction module. The second step, incremental pre-training stage: Based on the full corpus data in the field of airport operations, combined with the first, second and third level features output by the multi-level deep feature extraction module, the base model is incrementally pre-trained to enable the model to learn the professional knowledge, terminology system, business rules, control instruction logic and safety specifications in the field of airport takeoff control and flight release. The third step, the instruction fine-tuning stage: Based on the labeled instruction fine-tuning dataset, the global deep feature vector output by the multi-level deep feature extraction module is used as the core input, and the instruction semantic parsing results, conflict detection results, conflict level determination, and conflict resolution schemes are used as the output targets. The model is fine-tuned using a supervised fine-tuning method. At the same time, reinforcement learning with human feedback is combined. Senior tower controllers and airport operations management personnel score and provide feedback on the conflict resolution schemes output by the model to optimize the compliance and executability of the conflict resolution schemes generated by the model. The fourth step is model validation and optimization: The trained model is validated based on the test set. Validation metrics include semantic understanding accuracy, entity recognition accuracy, scene classification accuracy, conflict detection precision, recall, F1 score, and compliance rate of conflict resolution solutions. The weights of the multi-level deep feature extraction module, model parameters, and dataset are optimized and iterated to address validation deviations until the model performance meets the high safety and real-time requirements of airport takeoff control and flight release, thus completing the construction of the dedicated large model.

[0037] The airport operation instruction semantic understanding module performs real-time instruction semantic parsing. For the three multi-entity operational instructions input in real time, the first is a voice-transcribed text of the takeoff control instruction issued by the air traffic control tower to the aircraft crew: "Flight MU5678, runway 36 is clear for takeoff, surface wind 120 degrees 3 meters per second, corrected sea pressure 1013 hPa, proceed to the runway for takeoff immediately"; the second is a release instruction issued by the airline's operations control center to the flight crew: "Flight MU5678, due to the destination airport's weather standards not meeting the landing requirements, this flight's release permission is cancelled, takeoff is prohibited, taxi back to gate 4 and await further scheduling"; the third is a scheduling instruction issued by the airport operations command center to the apron control department: "Apron Control Department, after flight MU5678 takes off, gate 4 should be immediately allocated to flight HU7890 for parking." The airport operational instruction semantic understanding module executes the following processing flow: The first step is real-time instruction preprocessing: For the three real-time input control voice transcription texts, airline release instruction texts, and dispatch instruction texts mentioned above, perform cleaning and normalization preprocessing operations consistent with historical data to generate standardized texts; The second step, real-time extraction of third-level features: Through the trained multi-level deep feature extraction module, the standardized text is extracted step by step to generate the corresponding first-level basic semantic features, second-level scenario-based business logic features, and third-level global deep features. The third step is model reasoning and semantic parsing: The third-level global deep features are input into the trained airport operation-specific large language model. The model outputs standardized semantic parsing results, including the core element quadruple of the three instructions, business scenario, instruction type, instruction intent, compliance verification results, and entity standardization results. The business scenario of the first instruction is takeoff control, and the instruction intent is to permit the flight to take off immediately. The business scenario of the second instruction is flight release, and the instruction intent is to cancel the flight's takeoff permission and prohibit takeoff. The business scenario of the third instruction is resource scheduling, and the instruction intent is to allocate the corresponding gate after the flight takes off. The entity standardization results clearly show that the core execution object of the three instructions is flight MU5678, and the action instructions are mutually exclusive. The fourth step is semantic result output: the semantic parsing results are stored in a structured manner and pushed synchronously to the instruction conflict resolution module.

[0038] The instruction conflict resolution module performs full-process conflict handling and closed-loop management. The instruction conflict resolution module, based on semantic parsing results and global deep features, completes the detection, location, severity determination, and resolution scheme generation of instruction conflicts, realizing a closed loop for conflict handling. The specific process is as follows: The first step, conflict detection and localization: Based on the conflict-sensitive features in the third-level global deep features, combined with the semantic parsing results of the three instructions within the current 10-minute time window, instruction conflicts were detected through model inference; the conflict types were accurately located as action conflicts, responsibility conflicts, and condition conflicts, involving the three takeoff control instructions, flight release instructions, and gate scheduling instructions mentioned above, and the conflicting entities were the tower air traffic control department, the airline operations control center, the airport operations command center, and the flight MU5678 crew, the conflict point was the mutual exclusion of the takeoff clearance actions of flight MU5678, and the scope of impact was the takeoff order of runway 36, the operational safety of flight MU5678, the gate resource scheduling plan, and the normal parking of subsequent flights; The second step is to assess the conflict level and risk: Based on the pre-built airport operation safety risk level system, and combined with the conflict type, scope of impact, and urgency, the conflict level is determined to be extremely high. The potential safety risks of the conflict are flight safety hazards caused by flights taking off without valid clearance, runway operation conflicts, and operational impacts such as chaotic multi-flight gate scheduling, large-scale adjustments to flight operation plans, and passenger travel delays. The third step is to generate conflict resolution solutions: In response to this extremely high-risk conflict, based on civil aviation regulations, airport operation specifications, air traffic management rules, business rules, the current status of airport runway and parking space resources, and the weather conditions at the destination airport, two compliant conflict resolution solutions are generated using a domain-specific large language model. Each solution includes handling steps, responsible parties, execution time windows, resource allocation suggestions, and risk prevention and control measures. The fourth step is solution selection and recommendation: Based on preset decision indicators, the two solutions are ranked and optimized. The decision indicators are ranked from highest to lowest priority as follows: highest safety priority, least operational impact, highest execution efficiency, and least resource consumption. The optimal solution is as follows: The air traffic control tower immediately issues a termination order to the flight MU5678 crew, confirming that the flight has not entered the runway and canceling the takeoff clearance; the airline's operations control center simultaneously reports the reason for the flight cancellation and the subsequent scheduling plan to the airport operations command center and the air traffic control tower; the airport operations command center immediately adjusts the gate scheduling plan. The pre-assignment of gate 4 to flight HU7890 was cancelled, and gate 4 was reserved for flight MU5678 to taxi back and park. At the same time, a backup gate was reassigned to flight HU7890. Dedicated personnel from the airport operations command center and the tower control department were arranged to complete the collaborative verification of multi-entity instructions to ensure the consistency of subsequent instructions. Conflict information, risk level, and optimal resolution plan were simultaneously pushed to the tower air traffic control department, airline operations control center, airport operations command center, apron control department, and flight MU5678 crew. At the same time, the extremely high risk level emergency alarm of the visualization and emergency alarm module was triggered. The fifth step is to handle the closed loop and iterate the model: track the implementation of the conflict resolution plan throughout the process, record the handling results and feedback information from each party, and after confirming that the conflict has been completely resolved, there are no flight safety hazards or large-scale operational impacts, supplement the training dataset with the extremely high-risk case data of this handling, and immediately make targeted incremental fine-tuning to the domain large language model to enhance the detection and resolution capabilities of similar conflicts.

[0039] Visualization and emergency alarm modules are used in conjunction with each other. The visualization and emergency alarm module provides a full-process visualization of the semantic parsing results, conflict detection results, and resolution plan execution progress of this multi-subject command. It triggers corresponding emergency audible and visual alarms for conflicts of extremely high risk level, and supports traceable management of the entire command processing process and real-time monitoring of model effects.

[0040] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An airport operation semantic understanding and instruction conflict resolution system based on a large language model, characterized in that, include: The module includes a data preprocessing module, a multi-level deep feature extraction module, a domain-specific large language model training and fine-tuning module, an airport operation instruction semantic understanding module, an instruction conflict resolution module, and an airport operation domain knowledge base. Data preprocessing module: used to aggregate, clean, normalize, label and divide the airport operation historical instruction data from multiple sources, and output a standardized text dataset; Multi-level deep feature extraction module: used for third-level progressive deep feature extraction of standardized text data, wherein... The second-level feature extraction takes the output of the first-level feature extraction as its sole input, and the third-level feature extraction takes the output of the second-level feature extraction as its sole input. The final output is a global deep feature vector used for model training and inference. Domain-specific large language model training and fine-tuning module: used to perform incremental pre-training and instruction fine-tuning of pre-trained large language models based on global deep feature vectors, and to build an end-to-end large model for airport operation semantic understanding and instruction conflict resolution; Airport Operation Instruction Semantic Understanding Module: This module is used to preprocess real-time input airport operation instructions, extract third-level features, and perform model reasoning, outputting structured semantic parsing results. Instruction conflict resolution module: Based on semantic parsing results and global deep features, it completes the detection, location, level determination and resolution scheme generation of instruction conflicts, realizing a closed loop of conflict handling; Airport Operations Knowledge Base: Provides fundamental support for the entire process by including civil aviation regulations, airport operation standards, professional terminology, business rules, and conflict type systems.

2. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The specific processing flow of the data preprocessing module includes: Multi-source data aggregation: Aggregates historical airport operation data, including tower control voice transcription text, airport operation command center dispatch instruction text, ground support work order text, airline release instruction text, apron operation walkie-talkie call transcription text, emergency response instruction text, flight dynamic data, and airport resource status data; Data cleaning: Remove invalid data, including duplicate text, meaningless interjections, noisy characters, incomplete sentences, and filter text content that is irrelevant to airport operations; Text normalization: Standardize and unify the format of technical terms, flight numbers, gate numbers, runway numbers, taxiway numbers, time expressions, and spatial location expressions; Data annotation: Based on airport operation domain standards and business rules, text data is annotated with domain entities, scene tags, instruction types, and conflict tags; Dataset partitioning: The labeled dataset is divided into training set, validation set, and test set according to a preset ratio.

3. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The multi-level deep feature extraction module includes a first-level feature extraction unit, a second-level feature extraction unit, and a third-level feature extraction unit cascaded in sequence. The first-level feature extraction unit is used to extract basic semantic entities and syntactic features in the airport operation domain. The input is preprocessed standardized text, and the output is the first-level basic semantic feature vector matrix. The second-level feature extraction unit is used to perform airport operation scenario-based business logic feature extraction, and the output is the second-level scenario-based business logic feature vector matrix; The third-level feature extraction unit is used to perform cross-subject instruction interaction global dependency and conflict-sensitive feature extraction, and the output is a third-level global deep feature vector.

4. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The basic semantic entity and syntactic feature extraction performed by the first-level feature extraction unit specifically includes the following steps: Step 1: Domain Terminology and Entity Embedding Encoding: Based on the airport operation domain knowledge base, an airport-specific terminology lexicon and entity dictionary are constructed. The entity types in the entity dictionary include subject, resource, action, spatiotemporal constraint, and state types. An extended embedding is performed using a pre-trained large language model's token embedding layer combined with the domain terminology lexicon. After tokenizing the input text, an initial token embedding vector is generated. The named entity recognition sub-model is used to identify and classify airport domain entities in the text. Entity type masks are added to the identified entities to generate entity-aware embedding vectors. The initial token embedding vector and the entity-aware embedding vector are concatenated to obtain entity-enhanced word-level embedding features. The second step, syntactic dependency and sentence structure feature encoding, involves using a syntactic dependency analysis model to perform syntactic parsing on the input text, extracting subject-verb-object structures, modifier relationships, core predicates, and syntactic relationships between the initiator and receiver of the instruction, and generating a syntactic dependency feature matrix. Based on a pre-built airport instruction sentence structure template library, the sentence structure types of the input text are matched and classified to generate sentence structure type encoding vectors. The syntactic dependency feature matrix and the sentence structure type encoding vectors are then fused to obtain syntactic-level feature vectors. The third step, first-level feature fusion and normalization, involves cross-dimensional fusion of entity-enhanced word-level embedding features and syntactic-level feature vectors. A multi-head attention mechanism is used to assign weights to the fused features, strengthening the feature weights of core entities and core syntactic structures while weakening the weights of meaningless modifiers. Through layer normalization and linear transformation, a fixed-dimensional first-level basic semantic feature vector matrix is ​​generated.

5. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The scenario-based business logic feature extraction performed by the second-level feature extraction unit specifically includes the following steps: Step 1: Business Scenario Classification and Scenario Feature Anchoring: A scenario classification system is constructed based on the entire business process of airport operations. The scenario classification system includes eight major categories of core scenarios: takeoff control, landing control, ground taxiing, apron support, passenger service, emergency dispatch, flight release, and resource dispatch. Each major category is further subdivided into corresponding sub-scenarios. Based on the first-level basic semantic feature vector, a scenario classification sub-model is used to classify the business scenarios of input instructions, and the scenario classification results and scenario probability distribution vectors are output. Based on the pre-built business rule library for each scenario, the rule embedding vectors of the core business elements, constraint rules, and rights and responsibilities boundaries of the corresponding scenario are extracted. The scenario probability distribution vector and the rule embedding vector are concatenated to generate a scenario-aware rule feature vector. The second step, core business element extraction and association modeling of instructions: Based on the domain entities and syntactic structures identified by the first-level feature recognition, and combined with the rule feature vectors of scene awareness, the six core business elements of a single instruction are extracted: instruction initiating subject, instruction receiving subject, instruction execution action, execution object, spatiotemporal constraints, and execution conditions. A "subject-action-object-constraint" business element quadruple is constructed. A graph neural network is used, with each business element as a node and the business relationship between elements as an edge, to perform association modeling on the business element quadruple, generating business element association graph features. Through the business rule verification sub-model, the business compliance between elements is initially verified, and compliance feature codes are generated. The business element association graph features and compliance feature codes are fused to obtain the deep features of the business elements. The third step, instruction intent and type deep encoding: Based on the first-level sentence structure features and business element deep features, the core intent and instruction type of the instruction are classified and encoded to generate instruction intent encoding vector and instruction type encoding vector. These are then concatenated with the business element deep features to obtain instruction-level business semantic features. Step 4, Second-level feature fusion and dimensional alignment: The scene-aware rule feature vector, business element deep features, and instruction-level business semantic features are fused in multiple dimensions. A cross-attention mechanism is used to achieve deep interaction between the first-level basic semantic features and the second-level business logic features, strengthen the feature weights that conform to the current scene business rules, and filter out noisy features that are irrelevant to the business logic. Through layer normalization and linear transformation, a second-level scene-based business logic feature vector matrix that is aligned with the dimensions of the first-level feature vector is generated.

6. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The cross-subject instruction interaction global dependency and conflict-sensitive feature extraction performed by the third-level feature extraction unit specifically includes the following steps: Step 1: Modeling the timing and subject grouping features of instructions: For all instructions within the same preset time window, sort them according to their timestamps to generate an instruction timing sequence; Based on the instruction initiating subject and receiving subject information in the second-level features, group the instructions into subjects to generate subject grouping features; Use a timing coding model to encode the timing position, time window, and subject affiliation of the instructions to generate a timing-subject-aware coding vector, which is concatenated with the input second-level scenario-based business logic feature vector matrix to obtain a timing-subject-enhanced multi-instruction feature matrix; Step 2: Cross-instruction global dependency modeling: For the multi-instruction feature matrix with temporal-subject enhancement, a multi-head cross-attention mechanism is used to construct a cross-instruction global dependency graph. Each instruction is a node, and the business relationships between instructions are edges. These business relationships include temporal dependencies, subject responsibility relationships, resource consumption relationships, execution condition dependencies, and action mutual exclusion relationships. A graph Transformer network is used to perform deep encoding on the global dependency graph to mine the implicit dependencies between instructions and generate a cross-instruction global dependency feature matrix. Step 3: Conflict-Sensitive Feature Anchoring and Enhancement: Based on a pre-built airport operation instruction conflict type library, conflict-sensitive sites are anchored in the global dependency feature matrix. The conflict type library includes spatiotemporal conflicts, action conflicts, responsibility conflicts, temporal conflicts, and conditional conflicts. A contrastive learning mechanism is adopted to compare the features of positive samples labeled with conflict and negative samples without conflict, learn the difference distribution between conflict features and non-conflict features, strengthen the feature weights of conflict-sensitive sites, and weaken features unrelated to conflict. Conflict level encoding vectors are generated for different levels of conflict and fused with the anchored conflict-sensitive features to obtain global features enhanced by conflict. Step 4, Third-level feature aggregation and global feature output: Deeply fuse the multi-instruction features enhanced by temporal-subject, the global dependency features across instructions, and the global features enhanced by conflict. Use gated recurrent units to aggregate the temporal sequence features. Combining global average pooling and max pooling, the fused features are compressed and aggregated in dimension. After layer normalization and linear transformation, a fixed-dimensional third-level global deep feature vector is generated.

7. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The specific processing flow of the domain-specific large language model training and fine-tuning module includes: Base model selection and initialization: A general pre-trained large language model is selected as the base model, its pre-trained weights are loaded, and the embedding layer, attention layer and output layer of the model are adapted to make it suitable for the feature vector dimension output by the multi-level deep feature extraction module. Incremental pre-training stage: Based on the full corpus data of airport operations, combined with the first, second and third level features output by the multi-level deep feature extraction module, the base model is incrementally pre-trained to enable the model to learn the professional knowledge, terminology system, business rules and semantic logic of airport operations. Instruction fine-tuning stage: Based on the labeled instruction fine-tuning dataset, the global deep feature vector output by the multi-level deep feature extraction module is used as the core input, and the instruction semantic parsing results, conflict detection results, conflict level determination, and conflict resolution schemes are used as the output targets. The model is fine-tuned in a supervised manner, and the conflict resolution scheme generation ability of the model is optimized by combining human feedback reinforcement learning. Model Validation and Optimization: The trained model is validated based on the test set. Validation metrics include semantic understanding accuracy, entity recognition accuracy, scene classification accuracy, conflict detection precision, recall, F1 score, and compliance rate of conflict resolution solutions. The weights of the multi-level deep feature extraction module, model parameters, and dataset are optimized and iterated to address validation deviations until the model performance meets the safety and efficiency requirements of airport operations.

8. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The specific processing flow of the airport operation instruction semantic understanding module includes: Real-time instruction preprocessing: Perform preprocessing operations consistent with historical data on real-time input speech-to-text, dispatch instruction text, work order text, and walkie-talkie conversation text to generate standardized text; Real-time extraction of third-level features: Through the multi-level deep feature extraction module, the standardized text is extracted step by step to generate the corresponding first-level basic semantic features, second-level scenario-based business logic features, and third-level global deep features; Model reasoning and semantic parsing: Input the third-level global deep features into the trained domain language model, and the model outputs standardized semantic parsing results, including the instruction core element quadruple, business scenario, instruction type, instruction intent, compliance verification results, and entity standardization results; Semantic result output: The semantic parsing results are stored in a structured manner and pushed synchronously to the instruction conflict resolution module.

9. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The specific processing flow of the instruction conflict resolution module includes: Conflict detection and localization: Based on the conflict-sensitive features in the third-level global deep features, combined with the semantic parsing results of all instructions within the current time window, instruction conflicts are detected through model inference; if a conflict exists, the type of conflict, the instructions involved, the conflict subject, the conflict location, and the scope of impact are accurately located. Conflict Level and Risk Assessment: Based on a pre-built airport operation safety risk level system, and combined with the conflict type, scope of impact, and urgency, the conflict is classified into four levels: low risk, medium risk, high risk, and extremely high risk. The potential safety risks and operational impacts of the conflict are assessed simultaneously. Conflict resolution solution generation: For conflicts of different levels, multiple compliant conflict resolution solutions are generated by combining civil aviation regulations, airport operation standards, business rules, and current airport resource status through a domain-wide big language model. Each solution includes handling steps, responsible parties, execution time windows, resource allocation suggestions, and risk prevention and control measures. Solution optimization and push: Based on preset decision indicators, multiple resolution solutions are ranked and optimized. The decision indicators are ranked from high to low priority as follows: highest safety priority, least operational impact, highest execution efficiency, and least resource consumption. The optimal solution is output. The conflict information, risk level, and optimal resolution solution are simultaneously pushed to the corresponding responsible parties and the airport operation command center. Closed-loop management and model iteration: Track the implementation of conflict resolution solutions, record the results and feedback information, supplement the training dataset with the case data of completed cases, and perform incremental fine-tuning of the model on a regular basis.

10. The airport operation semantic understanding and instruction conflict resolution system based on a large language model according to claim 1, characterized in that: The airport operations knowledge base includes a civil aviation regulations database, an airport operation specifications database, a professional terminology dictionary, a business rules database, a scenario classification system, a conflict type database, an emergency plan database, and a historical handling case database, supporting real-time updates and iterations. The system also includes a visualization and emergency alarm module, used to visualize semantic parsing results, conflict detection results, and resolution solutions, triggering corresponding alarms for different levels of conflict, and supporting full-process instruction traceability and model effect monitoring.

Images