Flight special situation control instruction generation and state closed loop control method and system based on large language model

Abstract

The application relates to the technical field of flight simulation training and intelligent air traffic control, and discloses a flight special situation control instruction generation and state closed-loop control method and system based on a large language model, which comprises the following steps: extracting unified data records from flight special situation data, cutting and verifying to form audited small scenes, and constructing a special situation small scene library and a vector index; based on a training state source, a previous control instruction, a dialogue summary and a completed step, forming a current state set, performing safety hard constraint filtering and vector similarity matching on the special situation small scene library, and obtaining a candidate small scene set; combining candidate small scene constraint information and the current state set to form a model input package, calling a large language model to generate and verify a control instruction package; comparing a student recitation slot and real-time aircraft parameters with the control instruction package to form an observation result package, and performing state transition based on the observation result package. The application can improve the safety, accuracy and controllability of flight special situation control instruction generation, and realizes closed-loop promotion of training states.

Description

Technical Field

[0001] This invention relates to the fields of flight simulation training and intelligent air traffic control technology, specifically to a method and system for generating flight emergency control instructions and implementing closed-loop state control based on a large language model. Background Technology

[0002] Existing flight simulation training and air traffic control voice interaction systems typically generate training turns using fixed scripts, rule-based triggers, or a combination of conventional speech recognition and intent recognition. While these methods can handle predefined routine communication scenarios, in flight emergency training, emergency handling often involves multiple types of status information, including flight phase, anomaly type, real-time aircraft status, pilot requests, preceding instructions, historical dialogues, and completed handling steps. Furthermore, different states have strong temporal dependencies and safety constraints.

[0003] Existing scenario or rule bases typically organize data according to complete events, training subjects, or coarse-grained stages, making them difficult to directly serve as the smallest unit of action for matching in multi-turn dialogues. When trainees' recitations are incomplete or incorrect, aircraft status deviates from expectations, or runway or airspace status changes, fixed scripts struggle to adjust the next round of control dialogue in a timely manner, and single-field triggering rules are also insufficient to cover unforeseen combinations of states.

[0004] When relying solely on a large language model to freely generate control instructions, the model output may lack traceable instruction objectives, necessary repetitions, prohibited actions, and state transition condition constraints, resulting in inconsistencies between the generated content and the current aircraft status, operating rules, or established handling procedures.

[0005] The existing system still has difficulty reliably writing back the trainee's recitation verification results and the aircraft's real-time status verification results into the dialogue control process, which in turn affects the continuous progress of flight emergency training dialogue and the closed-loop control of the status.

[0006] Therefore, this application provides a method and system for generating flight emergency control instructions and implementing closed-loop state control based on a large language model to solve the above-mentioned technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for generating flight emergency control instructions and implementing closed-loop state control based on a large language model, in order to solve the technical problems in the prior art, such as the difficulty in matching complete event-level data round by round in flight emergency training, the difficulty in unifying heterogeneous states for retrieval, the lack of structured constraints in instructions generated by the large language model, and the difficulty in writing back the repetition and aircraft state verification results to the state control process.

[0008] To address the aforementioned technical problems, this invention provides a method for generating flight emergency control instructions and implementing state closed-loop control based on a large language model, comprising: Extract unified data records from flight emergency data, segment and verify the unified data records for processing steps to form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index; The training state source, preceding control instructions, dialogue summary, and completed steps are input into the hierarchical finite state machine to form the current state set with state version numbers; Perform hard-constraint filtering on the special scenario library based on the current state set, and perform vector similarity matching based on vector index to output a candidate scenario set; The constraint information in the candidate small scene set is converted into read-only constraint fields, which are combined with the current state set to form a model input package. Based on the model input package, the large language model is called to generate a control instruction package. After passing the safety constraint and state transition verification, the instruction text is sent. The trainee repeats the slot and the aircraft's real-time parameters, respectively, and compares them with the control instruction package to form an observation result package. Based on the observation result package, the hierarchical finite state machine is controlled to perform state transitions, forming an updated current state set or outputting a manual takeover flag.

[0009] In some specific embodiments, extracting a unified data record from flight incident data further includes: The system performs noise reduction, voice activity detection, speaker separation, or vocal tract annotation on the call audio, and uses an automatic speech recognition model adapted to the civil aviation communication corpus to generate transcribed text with start and end times and confidence scores. Match call sign, frequency, turn order, and speaker characteristics with the transcribed text to label controller or pilot roles. Call signs, Chinese numerals, units, and civil aviation terminology are standardized, and low-confidence transcribed texts are sent to a manual sampling queue. The output is a unified data record containing data identifier, source location, speaker role, start and end time, standardized text, confidence level, data type, and version.

[0010] In some specific embodiments, the unified data record processing steps are segmented and verified to form audited sub-scenes, which further include: The conversation turns are segmented based on speaker changes, silence intervals, and timestamps. The special situation type, applicable flight phase, pilot request, instruction target, permitted actions, prohibited actions, necessary repeats, state transition conditions, and completion conditions are extracted from the conversation turns, report text, debriefing text, and operation rule text. The extracted fields within the same processing step are combined into candidate sub-scenes. The enumeration values, units, numerical ranges, dependencies, and source integrity are verified. After verification and review, the approved sub-scenes are output.

[0011] In some specific embodiments, the structured constraint information of the approved minor scenarios is written into the special incident minor scenario library, and the corresponding scenario vector is written into the vector index, further including: Store state constraints, necessary repetitions, and state transition conditions in a key-value structure; store source extracts and example expressions in text format; and connect adjacent processing steps using predecessor and successor identifiers. The special situation type, applicable flight phase, state constraints, pilot request, preceding actions and handling objectives are serialized into scene retrieval text, and then input into the text embedding model to obtain scene vectors. Write the scene vector along with the sub-scene identifier, document version, review version, and embedded model version into the vector index.

[0012] In some specific embodiments, the training state source, preceding control instructions, dialogue summary, and completed steps are input into a hierarchical finite state machine to form a current state set with state version numbers, further including: Obtain the initial parameters of the training course or the previous round of updated status, the simulator's real-time flight parameters, pilot requests, preceding control instructions, summaries of the most recent rounds of dialogue, and completed steps, and input them into a hierarchical finite state machine; Constrained by the transition conditions of the top-level emergency process status and emergency handling sub-states, the flight phase, emergency type, severity, airspace or runway, flight altitude, flight speed, flight heading, equipment status, weather conditions, conflict target, current handling target, remaining handling time, and data confidence are written into the current state set.

[0013] In some specific embodiments, the special situation mini-scenario library is subjected to safety hard constraint filtering based on the current state set, and vector similarity matching is performed based on vector indexes to output a candidate mini-scenario set, further including: Using special situation type, flight phase, equipment availability, prohibited actions, safety boundaries and preceding steps as hard constraints, retain small scenarios that meet the conditions of enumeration inclusion, continuous quantity tolerance, timing sequence and duration, and eliminate small scenarios that do not meet the conditions of enumeration inclusion, continuous quantity tolerance, timing sequence and duration. The current status set, pilot request, preceding control instructions, dialogue summary, completed steps and current action target are serialized into query text and converted into query vectors. The cosine similarity between the query vector and the filtered scene vector is calculated, and the cosine similarity is used as the semantic similarity. The filtered small scenes are then weighted and sorted based on semantic similarity, key field matching degree, process continuity and data credibility, and a preset number of candidate small scenes with a value higher than the minimum acceptance threshold are output. If no qualified candidates are found, expand the scope of non-critical searches while retaining hard safety constraints. If no candidates are found, output a manual takeover flag.

[0014] In some specific embodiments, the constraint information in the candidate small scene set is converted into read-only constraint fields, which are then combined with the current state set to form a model input packet. Based on the model input packet, a large language model is called to generate a control instruction packet. After passing the safety constraint and state transition verification, the instruction text is sent. Further, this includes: Select at least one candidate scenario from the candidate scenario set that has the same source version and whose state constraints do not conflict. Write the scenario identifier, source identifier, instruction target, allowed action, prohibited action, necessary repetition, state transition condition and completion condition into the read-only constraint field. The model input package is composed of read-only constraint fields, current state set, system-level control rules, prohibited items, and structured output patterns, and is used as input for the large language model to generate control instruction packages; The large language model is called based on read-only constraint fields to generate control instruction packages. Field inheritance consistency checks, security boundary checks, and state reachability checks are performed on the control instruction packages. After the checks pass, the instruction text is sent.

[0015] In some specific embodiments, the trainee repeats the slot and the aircraft's real-time parameters are compared with the control instruction packet to form an observation result packet. Based on the observation result packet, the hierarchical finite state machine is controlled to perform state transitions, forming an updated current state set or outputting a manual takeover flag. This further includes: After sending the instruction text, the student's voice or keyboard text is standardized and the call sign, altitude, heading, speed, runway, frequency, permitted action and restriction condition slots are extracted; The key slots are compared with the necessary repetition items for precise values ​​or configuration tolerances. The non-key expressions are compared with the instruction text in the control instruction package for semantic similarity. The real-time parameters of the aircraft are compared with the expected state transition conditions and the fields to be observed in the next round, forming an observation result package containing the repetition verification results and the aircraft state satisfaction. When the repetition is completely correct and the aircraft state meets the expected state transition conditions, the hierarchical finite state machine controls the next disposal sub-state. When the recitation is correct and there are no key item conflicts, the hierarchical finite state machine is controlled to maintain the current processing sub-state and generate missing item prompts. When there is a repetition error, a critical item conflict, or the aircraft state deviates from the expected state, the hierarchical finite state machine controls the generation of corrective requirements or the re-retrieval of requirements. Output a manual takeover flag when there are no qualified candidates, continuous low confidence, continuous generation failures, or when a security boundary is triggered.

[0016] In some specific embodiments, when the recitation is correct and there are no key item conflicts, the hierarchical finite state machine is controlled to maintain the current processing sub-state and generate a missing item prompt requirement; when the recitation is incorrect, there is a key item conflict, or the aircraft state deviates from expectations, the hierarchical finite state machine is controlled to generate a correction requirement or a re-retrieval requirement, further including: Write the repetition error slots, non-critical missing expressions, and aircraft status deviation fields into the anomaly observation queue; The anomaly observation queue is read using critical slot conflicts, missing non-critical expressions, and recoverable deviations in aircraft status as anomaly categories. The system generates correction instructions for critical slot conflicts, missing item prompts for non-critical missing expressions, and re-retrieval requests for recoverable deviations in aircraft status. Write the generated requirement categories into the updated current state set.

[0017] Based on the same concept, the present invention also provides a flight emergency control instruction generation and state closed-loop control system based on a large language model, comprising: The small scenario library construction module is configured to extract unified data records from flight emergency data, perform processing steps and verification on the unified data records to form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index. The current state formation module is configured to input the training state source, preceding control instructions, dialogue summary and completed steps into a hierarchical finite state machine to form a current state set with state version numbers; The candidate small scene retrieval module is configured to perform safety hard constraint filtering on the special situation small scene library based on the current state set, and perform vector similarity matching based on the vector index to output the candidate small scene set; The control instruction constraint generation and verification module is configured to convert the constraint information in the candidate small scene set into read-only constraint fields, form a model input package with the current state set, call the large language model based on the model input package to generate a control instruction package, and send the instruction text after passing the safety constraint and state transition verification. The state-closed-loop control module is configured to compare the trainee's repeating slot and the aircraft's real-time parameters with the control instruction package to form an observation result package. Based on this observation result package, it controls a hierarchical finite state machine to execute state transitions, generating an updated current state set or outputting a manual takeover flag. Compared to existing technologies, its advantages are: This invention discloses a method and system for generating flight emergency control instructions and implementing closed-loop state control based on a large language model. The method decomposes flight emergency data into approved mini-scenes that include the handling objective, control actions, necessary repetition items, state constraints, and state transition conditions. The structured constraint information of the approved mini-scenes is written into an emergency mini-scene library, and the corresponding scene vectors are written into a vector index. Complete event-level data can be converted into handling step-level units that are searchable, traceable, and verifiable in each round. Each round of control dialogue can be associated with specific data sources and handling boundaries.

[0018] The training state source, preceding control instructions, dialogue summaries, and completed steps are input into a hierarchical finite state machine to form a current state set with state version numbers. Flight phases, situation types, real-time aircraft parameters, pilot requests, preceding control instructions, dialogue summaries, and completed steps can all enter the same structured state snapshot, and online retrieval can continuously advance around the current training state.

[0019] After applying hard safety constraints to the special scenario library based on the current state set, vector similarity matching is performed based on the vector index. Hard constraints such as device availability, prohibited actions, safety boundaries, and predecessor steps can work together with semantic matching. The candidate scenario set can simultaneously reflect the constraints of operating rules, the continuity of processes, and the semantic matching relationship.

[0020] The instruction targets, permitted actions, prohibited actions, necessary repetitions, state transition conditions, and completion conditions in the candidate small scene set are converted into read-only constraint fields and combined with the current state set to form the model input package. When the large language model generates the control instruction package, it has a clear business input range. The candidate small scenes and the current state set can jointly constrain the generation of instruction text, control actions, necessary repetitions, and state transition conditions.

[0021] The system performs field inheritance consistency checks, security boundary checks, and state reachability checks on the control instruction packets. If the checks fail, the erroneous fields are combined with the original model input packets to form a regenerated input. The model output is sent after independent verification. This allows the system to identify and intercept issues such as missing fields, inconsistent sources, prohibited actions, flight conflicts, and unreachable state transitions before the instructions are sent. If necessary, the system will initiate a manual takeover process.

[0022] The trainee repeats the slot and the aircraft's real-time parameters, respectively, and compares them with the control instruction package to form an observation result package. Based on the observation result package, the hierarchical finite state machine is controlled to perform state transitions. The repeat verification results and the aircraft state verification results can be stably written back to the state control process. Multi-round flight emergency control training can form a closed loop between advance, hold, back, re-retrieval, manual takeover and termination. Attached Figure Description

[0023] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating some specific embodiments of the flight emergency control instruction generation and state closed-loop control method based on a large language model according to the present invention. Figure 2 This is a schematic diagram of the structure of a flight emergency control command generation and state closed-loop control system based on a large language model according to some specific embodiments of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms "a," "the," and "the" as used in the embodiments of this application are also intended to include the plural forms, unless the context clearly indicates otherwise, and "multiple" generally includes at least two.

[0026] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0027] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0028] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0029] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0030] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0031] Reference Figure 1 A method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model, comprising: S101: Extract unified data records from flight emergency data, segment and verify the unified data records for processing steps, form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index. S102, input the training state source, preceding control instructions, dialogue summary and completed steps into the hierarchical finite state machine to form the current state set with state version number; S103: Perform hard safety constraints filtering on the special situation scenario library based on the current state set, and perform vector similarity matching based on vector index to output the candidate scenario set; S104: Convert the constraint information in the candidate small scene set into read-only constraint fields, and form a model input package with the current state set. Based on the model input package, call the large language model to generate a control instruction package. After passing the safety constraint and state transition verification, send the instruction text. S105 compares the trainee's repeat slot and the aircraft's real-time parameters with the control instruction package to form an observation result package. Based on the observation result package, it controls the hierarchical finite state machine to perform state transitions, forming an updated current state set or outputting a manual takeover flag.

[0032] To provide a clearer explanation, the steps in the embodiments of the present invention are described in detail below: S101: Extract unified data records from flight emergency data, segment and verify the unified data records for processing steps, form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index. In this embodiment, flight emergency data includes publicly available safety reports, authorized flight emergency control communication audio, training debriefing text, and operational rule text. Flight emergency data enters the data access and preprocessing flow, and processing paths are assigned according to data type. Communication audio enters the speech-to-text path, while report text, training debriefing text, and operational rule text enter the text parsing path. Each piece of data is assigned a data identifier, and the data source, data version, rule version, access time, and data segment position are recorded, enabling subsequent generated approved mini-scenes to be traced back to the data segment.

[0033] Furthermore, noise reduction, voice activity detection, speaker separation or vocal tract annotation are performed on the call audio, and an automatic speech recognition model adapted to the civil aviation communication corpus is used to generate transcribed text with start and end times and confidence scores. For call audio, the audio files are formatted uniformly according to sampling rate, number of channels, and volume range. Noise reduction processing is used to remove continuous background noise, short-term impulse noise, and equipment current noise. Speech activity detection is used to segment continuous audio into speech segments and silence segments. Speaker separation or channel labeling is used to distinguish between controller speech segments and pilot speech segments. When the audio is recorded in stereo, role initialization is prioritized based on the channel source. When the audio is recorded in mono, role initialization is performed by combining voiceprint clustering, call sign occurrence position, and turn order.

[0034] An automatic speech recognition model adapted from the civil aviation speech corpus was used to process the audio features of each speech segment, obtaining the transcribed text, the start time of the speech segment, the end time of the speech segment, and the confidence level. The automatic speech recognition model adapted from the civil aviation speech corpus can be an adaptation model that adds civil aviation call signs, runways, frequencies, altitudes, speeds, headings, and clearance action vocabulary to a general speech recognition model, or it can be a civil aviation-specific speech recognition model. For numbers, units, and control terms in the transcribed text, two fields are retained: the original transcription result and the normalized result. The original transcription result is used for manual review, and the normalized result is used for subsequent segmentation of smaller scenarios.

[0035] Match call sign, frequency, turn order, and speaker characteristics with the transcribed text to label controller or pilot roles. Call signs, frequencies, question-and-answer prompts, permission action words, and repeat action words are extracted from the transcribed text. Controller turns typically contain expressions of permission, instruction, restriction, or inquiry, while pilot turns typically contain expressions of call sign repeats, requests, confirmations, or reports. Call sign location, frequency broadcasting method, turn-related relationships, and speaker characteristics are used for role labeling. Role labeling output includes speaker role, role confidence, role basis field, and role verification status. When the same call sign appears in consecutive turns and the semantic relationship is "instruction-repetition," a turn association identifier is established between the controller's turn and the pilot's turn. When the role confidence of the same speech segment is lower than the exemplary threshold of 0.75, the speech segment is marked as a low-confidence speech segment and sent to the manual sampling queue.

[0036] Standardize call signs, Chinese numerals, units, and civil aviation terminology, and send low-confidence transcribed texts to a manual sampling queue. Output a unified data record containing data identifier, source location, speaker role, start and end time, standardized text, confidence level, data type, and version. Transcribed text is processed according to civil aviation communication dictionaries, call sign rules, frequency formats, number reading methods, and unit conversion rules. For example, "3,000 meters," "maintain 3,000 meters," and "altitude 3,000" can be standardized as an altitude field and altitude value; "turn right and head 180" can be standardized as a heading field and heading value 180; and "contact tower 120.3" can be standardized as a frequency field and frequency value 120.3. The standardized fields are written into a unified data record, which includes at least the data identifier, source location, speaker role, start and end times, standardized text, confidence level, data type, and version. Low-confidence transcribed text enters a manual review queue. Based on the manual review results, the low-confidence field is updated, while the original confidence level and review mark are retained. The unified data record serves as input for subsequent processing steps of segmentation and verification.

[0037] The conversation turns are segmented based on speaker changes, silence intervals, and timestamps. The special situation type, applicable flight phase, pilot request, instruction target, permitted actions, prohibited actions, necessary repeats, state transition conditions, and completion conditions are extracted from the conversation turns, report text, debriefing text, and operation rule text. Speaker changes are used to determine whether adjacent speech segments belong to different turns; silence intervals are used to determine whether consecutive speech segments from the same speaker need to be merged; and timestamps are used to determine the sequence of turns. For example, when adjacent speech segments have the same speaker role and the silence interval is less than 1.2 seconds, the adjacent speech segments are merged into the same turn; when adjacent speech segments have different speaker roles or the silence interval reaches 1.2 seconds, a new turn is formed. The report text, debriefing text, and operating rule text are segmented according to title, paragraph number, rule clause, action verb, and conditional verb.

[0038] Field extraction can be performed using rule parsers, information extraction models, large language model prompt word extraction, or manual annotation. The extraction process uses a single action target, a corresponding control action, an expected repetition, and a decidable state transition as the semantic boundary of a small scenario. From the data fragment, it extracts the situation type, applicable flight phase, pilot request, instruction target, permitted actions, prohibited actions, necessary repetitions, state transition conditions, and completion conditions. Necessary repetitions may include call sign, altitude, heading, speed, runway, frequency, permitted actions, and restrictions. State transition conditions may include the aircraft reaching the target altitude, heading entering the target heading tolerance, speed within the restricted range, runway status becoming clear, conflict target resolved, or equipment status restored.

[0039] The extracted fields within the same processing step are combined into candidate sub-scenes. The enumeration values, units, numerical ranges, dependencies, and source integrity are verified. After verification and review, the approved sub-scenes are output. Fields within the same handling step are grouped according to a small-scenario data structure. A candidate small scenario must at least include a scenario identifier, source identifier, situation type, applicable flight phase, predecessor identifier, successor identifier, state constraints, pilot request, instruction target, permitted actions, prohibited actions, necessary repetitions, state transition conditions, completion conditions, and example expressions. The scenario identifier is used to uniquely identify the small scenario, the source identifier is used to locate a unified data record or text data paragraph, and the predecessor and successor identifiers are used to describe the flow sequence between handling steps.

[0040] Perform field validation on candidate scenarios. Enumeration value validation checks whether the situation type, flight phase, permitted actions, and prohibited actions belong to a preset enumeration set; unit validation checks whether the altitude, speed, heading, frequency, and time fields use a unified unit; numerical range validation checks whether the altitude, speed, heading, frequency, and time fields fall within the allowed range of their respective fields; dependency validation checks whether the output of the preceding step can be used as the input of the current step; and source integrity validation checks whether the command targets, necessary repetitions, and state transition conditions in the candidate scenario all have a source. After successful validation and review, the candidate scenario is confirmed as an approved scenario; if validation fails, the candidate scenario is sent to the manual review queue, and the failed fields, reason for failure, source fragment, and review status are recorded.

[0041] Store state constraints, necessary repetitions, and state transition conditions in a key-value structure; store source extracts and example expressions in text format; and connect adjacent processing steps using predecessor and successor identifiers. State constraints, necessary repeats, and state transition conditions are stored using key-value structures or equivalent structured objects. State constraints can include situation type, applicable flight phase, flight altitude range, speed range, heading range, runway status, airspace status, equipment status, weather conditions, conflicting targets, and pilot requests. Necessary repeats record the slot name, expected value, tolerance, whether it is a critical item, and comparison method. State transition conditions record the aircraft status, pilot repeat status, and time conditions required to complete the current action step. Source excerpts and example expressions are stored in text format to provide traceable linguistic support when constructing the model input package.

[0042] The special situation type, applicable flight phase, state constraints, pilot request, preceding actions and handling objectives are serialized into scene retrieval text, and then input into the text embedding model to obtain scene vectors. The scenario retrieval text is constructed by concatenating the situation type, flight phase, state constraints, pilot request, preceding actions, and resolution objective in a fixed field order. For example, the scenario retrieval text could take the form of "Situation Type: Runway Occupied; Flight Phase: Approach; State Constraint: Aircraft has established an approach, runway unavailable; Pilot Request: Continue Approach; Preceding Action: Continue Approach Clearance; Resolution Objective: Go-Ahead Avoidance". The scenario retrieval text is then input into a text embedding model to obtain a scenario vector. The scenario vector can have dimensions of 384, 768, or 1024, consistent with the text embedding model used. The text embedding model can be a locally deployed model or a model invoked via a controlled interface; the model output is used for similarity calculation.

[0043] Write the scene vector along with the sub-scene identifier, document version, review version, and embedded model version into the vector index.

[0044] The vector index stores the mapping between scene vectors and sub-scene identifiers, while the special sub-scene library stores the structured fields of sub-scenes. It also records the document version, review version, and embedded model version, enabling online retrieval to identify the document version upon which a candidate sub-scene is based. When a sub-scene is added or modified, only the newly added or modified sub-scene has its scene vector regenerated and its vector index updated; when the embedded model version changes, scene vectors are regenerated for all sub-scenes and the vector index is rebuilt. Before reconstruction is complete, the published version of the sub-scene library and vector index continues to be used during the online training phase.

[0045] In this embodiment, a combination of manual and automatic transcription can be used to generate unified data records. The manually transcribed results are used as high-confidence text for role labeling, while the automatically transcribed results are used as supplementary fields for standardization. Small scene extraction can also employ a combination of rule extraction, information extraction models, and manual verification. When the model-extracted fields conflict with the rule parser's verification results, the source fragment is retained and sent to the manual review queue. After successful review, it is written into the special situation small scene library.

[0046] S102, input the training state source, preceding control instructions, dialogue summary and completed steps into the hierarchical finite state machine to form the current state set with state version number; In this embodiment, the training state source includes initial parameters of the training subject, the previous round of updated state, real-time flight parameters of the simulator, pilot requests, preceding control instructions, summaries of recent rounds of dialogue, completed steps, and the currently valid rule version. Real-time flight parameters of the simulator may include position, altitude, speed, heading, vertical position, runway or airspace status, equipment status, weather conditions, and conflict targets. Preceding control instructions are derived from the previous round of control instruction packets, summaries of recent rounds of dialogue are derived from the training record module, and completed steps are derived from the historical processing states of the hierarchical finite state machine.

[0047] Furthermore, the initial parameters of the training course or the previous round of updated status, the real-time flight parameters of the simulator, pilot requests, preceding control instructions, summaries of the most recent rounds of dialogue and completed steps are obtained and input into a hierarchical finite state machine; The hierarchical finite state machine comprises a top-level emergency procedure state and emergency handling sub-states. The top-level emergency procedure state can include emergency start, identification and confirmation, handling instructions pending generation, repetition pending verification, aircraft status pending verification, handling continuation, manual takeover, and emergency end. Emergency handling sub-states are refined according to the emergency type; for example, in a runway occupancy scenario, it could include continued approach monitoring, go-around instruction, go-around climb confirmation, and post-go-around trajectory handling. When initial parameters for the training exercise are input, an initial current state set is formed; when the previous round of updated states is input, the previous round's state version number, preceding control instructions, observation result packets, and state transition records are retained.

[0048] Constrained by the transition conditions of the top-level emergency process status and emergency handling sub-states, the flight phase, emergency type, severity, airspace or runway, flight altitude, flight speed, flight heading, equipment status, weather conditions, conflict target, current handling target, remaining handling time, and data confidence are written into the current state set.

[0049] The fields in the training state source are uniformly formed into a structured state snapshot. The flight phase can be derived from the initial parameters of the training subject or the real-time flight parameters of the simulator; the emergency type and severity can be derived from the training subject, instructor input, or identified emergency states; the airspace or runway comes from the simulator environment state or training environment settings; the flight altitude, flight speed, and flight heading come from the real-time parameters of the simulator; the equipment status and weather conditions come from the training state source; the conflict target comes from the traffic conflict detection results or training subject settings; the current handling target comes from the combination of the top-level emergency process status and the emergency handling sub-status; the remaining handling time is determined based on the current time, the emergency handling time limit, and the step start time; and the data confidence level is determined based on the field source, update time, and identification confidence level.

[0050] For example, data confidence can be formed by three components: source confidence, time freshness, and field consistency, according to configured weights. Source confidence comes from the confidence level of the simulator interface, automatic speech recognition results, or manual input; time freshness is determined based on the interval between the current time and the field update time, with a lower value for a longer interval; field consistency is determined based on the degree of consistency of the same field across different sources. Data confidence is used for filtering candidate small scenarios and judging consecutive low-confidence scenarios.

[0051] When a field is missing from the training state source, boundary processing is performed according to the field type. For simulator real-time fields such as flight altitude, flight speed, and flight heading, the most recent timestamped field value that is within the valid time window is read; when a usable field value is missing, the corresponding field is written to the low-confidence field. For dialogue fields such as pilot requests, preceding control instructions, and dialogue summaries, the current round or the records of the most recent rounds are read; when a usable dialogue field is missing, the current state set retains null values ​​and uses the null value fields for non-critical retrieval scope expansion processing. The state version number increments according to the number of state updates and is used along with the current state set for candidate small-scene retrieval and control instruction generation processing.

[0052] In this embodiment, a state diagram or a workflow engine capable of expressing equivalent transition conditions can also be used to form the current state set. The state diagram or workflow engine stores the top-level emergency process state, emergency handling sub-states, field sources, state version numbers, and transition records, and forms the current state set according to the same input fields. Under different implementation methods, the data structure of the current state set remains consistent so that subsequent retrieval, model input package construction, and state closure update are processed along the same fields.

[0053] S103: Perform hard safety constraints filtering on the special situation scenario library based on the current state set, and perform vector similarity matching based on vector index to output the candidate scenario set; In this embodiment, the candidate minor scenario retrieval receives a current state set, a minor scenario library, and a vector index. The current state set includes flight phase, minor scenario type, equipment status, runway or airspace status, flight altitude, flight speed, flight heading, pilot request, preceding air traffic control instructions, dialogue summary, completed steps, current action target, remaining action time, and data confidence level. The minor scenario library includes the state constraints, permitted actions, prohibited actions, necessary repetitions, state transition conditions, and completion conditions for approved minor scenarios. The vector index includes scenario vectors and minor scenario identifiers.

[0054] Furthermore, using special situation type, flight phase, equipment availability, prohibited actions, safety boundaries and preceding steps as hard constraints, small scenarios that meet the conditions of enumeration inclusion, continuous quantity tolerance, timing sequence and duration are retained, while small scenarios that do not meet the conditions of enumeration inclusion, continuous quantity tolerance, timing sequence and duration are eliminated. Hard constraint filtering for safety is used to handle safety constraints independent of semantic similarity. Situation types and flight phases are judged using equivalence or set inclusion criteria; equipment availability, runway status, and airspace status are judged according to state constraint fields; the prohibited actions fields of candidate sub-scenarios are compared item by item with the equipment status, runway status, airspace status, and conflicting targets in the current state set; safety boundaries are determined based on prohibited items in the operating rules and current flight parameters; and preceding steps are determined based on completed steps and the preceding identifiers of candidate sub-scenarios. Continuous quantity fields are judged using tolerance criteria; for example, altitude, speed, heading, and remaining handling time can be set with tolerances or ranges. Timing conditions are judged using sequence and duration criteria; for example, if a handling step requires a go-around instruction to have been issued and a preset observation time has elapsed, the interval between the instruction issuance time and the current time in the state transition record is checked.

[0055] The current status set, pilot request, preceding control instructions, dialogue summary, completed steps and current action target are serialized into query text and converted into query vectors. The query text is generated according to the same or compatible field order as the scenario retrieval text. The query text may include "Current Flight Phase," "Current Situation Type," "Status Constraint Fields," "Pilot Request," "Preceding Control Instructions," "Summaries of Recent Rounds of Dialogue," "Completed Steps," and "Current Action Objective." The query text is input into a text embedding model to obtain a query vector. If certain fields are in a low-confidence state, the field names and low-confidence markers are retained in the query text, and the contribution of low-confidence fields to the matching degree of key fields is reduced in subsequent scoring.

[0056] The cosine similarity between the query vector and the filtered scene vector is calculated, and the cosine similarity is used as the semantic similarity. The filtered small scenes are then weighted and sorted based on semantic similarity, key field matching degree, process continuity and data credibility, and a preset number of candidate small scenes with a value higher than the minimum acceptance threshold are output. The cosine similarity between the query vector and the filtered scene vector is calculated to represent the semantic similarity between the query text and the filtered scene.

[0057] Key field matching accuracy can be determined based on the number of hits and comparable data for situation type, flight phase, pilot request, preceding actions, current handling objective, and necessary state constraints. Process continuity can be determined based on whether the preceding identifier of the candidate sub-scenario falls into a completed step, whether the candidate sub-scenario is a permissible subsequent step of the current handling sub-state, and whether the candidate sub-scenario is connected to preceding control instructions. Data credibility can be determined based on data source type, review version, data version, and manual review status.

[0058] Ranking weights are assigned to semantic similarity, key field matching, process continuity, and data credibility. Based on these parameters and their corresponding ranking weights, the total candidate score for each filtered sub-scene is determined, with the sum of all ranking weights being 1. The filtered sub-scenes are then sorted from highest to lowest total score, and a preset number of candidate sub-scenes with total scores exceeding a minimum acceptable threshold are output. Each candidate sub-scene retains its scene identifier, source identifier, score, field match status, and version information.

[0059] For example, the minimum acceptance threshold can be set to 0.65, and the preset number can be set to 3 or 5. The above values ​​are exemplary values ​​in the implementation description and can be configured according to the actual training environment.

[0060] The prohibited actions in the candidate small scene are compared with the device status, airspace or runway status, and conflicting targets in the current state set one by one to generate hard conflict markers. Read the prohibited action field of the candidate sub-scene, as well as the device state, airspace or runway state, and conflict target in the current state set, and compare the prohibited action field with the corresponding state item by item. When the current device state, airspace or runway state, or conflict target corresponds to a prohibited action of the candidate sub-scene, generate the corresponding hard conflict flag for the candidate sub-scene.

[0061] For permitted actions in candidate sub-scenes, security checks can be performed based on the current device status, airspace or runway status, and conflicting targets. When a permitted action conflicts with the current device status, airspace or runway status, or a conflicting target, a corresponding hard conflict flag is generated. The hard conflict flag is written into the filtering record of candidate sub-scenes and used for the elimination or sorting of candidate sub-scenes.

[0062] Candidate sub-scenes with hard conflict markers are eliminated; supplementary ranking scores are calculated for candidate sub-scenes without hard conflict markers using the preceding step continuity, state constraint hit, and source review version as ranking parameters; the supplementary ranking scores are then incorporated into the scores of the candidate sub-scene set. Candidate sub-scenes marked with hard conflict are eliminated, while supplementary ranking scores are calculated for candidate sub-scenes without hard conflict marks. Precursor step continuity indicates that the predecessor identifier of the candidate sub-scene matches the end of the completed steps in the current state set; state constraint hit indicates that the current state set fields satisfy the state constraint fields of the candidate sub-scene; source review version indicates whether the corresponding document version and review version of the candidate sub-scene are in the current release version. Supplementary ranking scores are determined based on predecessor step continuity, state constraint hit, and source review version, and these supplementary ranking scores are incorporated into the score of the candidate sub-scene set.

[0063] If no qualified candidates are found, expand the scope of non-critical searches while retaining hard safety constraints. If no candidates are found, output a manual takeover flag.

[0064] No qualified candidates indicate that the filtered sub-scenes are empty, or the total score of all candidate sub-scenes is below the minimum acceptance threshold. While retaining hard safety constraints, the non-critical search scope is expanded. Expansion methods include relaxing non-critical semantic fields, increasing the time window for dialogue summaries, reducing the weight of pilot request fields, or expanding the set of synonymous expressions for the handling target. Hard safety constraints include prohibited actions, safety boundaries, equipment availability, runway or airspace status, and necessary conditions for preceding steps. After expanding the non-critical search scope, hard safety constraint filtering and vector similarity matching are re-executed. If no qualified candidates are found after the re-search, a manual takeover flag is output, and the reason for manual takeover is recorded as "no qualified candidates," initiating a status update process.

[0065] In this embodiment, keyword retrieval, rule scoring, and vector retrieval can also be combined for candidate small-scene retrieval. Keyword retrieval is used to match structured or semi-structured fields such as call sign, runway, frequency, flight phase, and emergency type; rule scoring is used to calculate safety hard constraints and process continuity; and vector retrieval is used to handle unconventional expressions and request texts with different word orders. The three types of results enter the same ranking framework, preserving the priority of safety hard constraints, and using similarity and field matching degree as the candidate ranking criteria.

[0066] S104: Convert the constraint information in the candidate small scene set into read-only constraint fields, and form a model input package with the current state set. Based on the model input package, call the large language model to generate a control instruction package. After passing the safety constraint and state transition verification, send the instruction text. In this embodiment, the control instruction generation process receives a set of candidate sub-scenes and a set of current states, and reads system-level control rules, prohibited items, and structured output patterns. The model input package is used to constrain the large language model to generate control instruction packages. The constraint information in the candidate sub-scene set is converted into read-only constraint fields in the model input package, and based on the read-only constraint fields, the current state set, and the structured output pattern, the large language model is invoked to generate control instruction packages that conform to the structured output pattern.

[0067] Furthermore, select at least one candidate small scene from the candidate small scene set that has the same source version and whose state constraints do not conflict. Check the document version, approval version, and embedded model version of the candidate scenarios. Consistent source versions indicate that the document and approval versions used by the candidate scenarios are currently released or permitted versions. No conflicting state constraints mean that there are no conflicts between multiple candidate scenarios regarding flight phases, runway states, airspace states, equipment states, permitted and prohibited actions, or predecessor / successor order. If the state constraints of the highest-scoring candidate scenario and the second-highest-scoring candidate scenario do not conflict, multiple candidate scenarios can be selected simultaneously and written into the model input package in score order. If there are state constraint conflicts among candidate scenarios, retain the scenario with the higher score and that meets the hard safety constraints.

[0068] Write the scene identifier, source identifier, instruction target, allowed actions, prohibited actions, necessary repetitions, state transition conditions, and completion conditions into a read-only constraint field; Read-only constraint fields are stored as structured objects. Scene identifiers are used to locate candidate sub-scenes, source identifiers are used to locate data sources, instruction targets are used to limit the disposal objectives to be achieved in this round of control instructions, permitted actions are used to limit the selectable control actions, prohibited actions are used to restrict generated content, necessary repeats are used to form repeat verification fields in the control instruction package, state transition conditions are used to form expected state transition conditions, and completion conditions are used to determine whether the special situation handling has entered the final state. Read-only constraint fields can be set as input fields that cannot be modified by the large language model. Field inheritance consistency checks are used to verify whether the output of the large language model uses read-only constraint fields.

[0069] The model input package is composed of read-only constraint fields, current state set, system-level control rules, prohibited items, and structured output patterns, and is used as input for the large language model to generate control instruction packages; The model input package includes the current state set, read-only constraint fields for candidate sub-scenes, system-level control rules, prohibited items, output field definitions, version information, and the current round input package identifier. The structured output mode requires that the large language model output control instruction packages must include a scene identifier, instruction text, control action, necessary repetitions, state transition conditions, fields to be observed in the next round, source identifier, and confidence level. The structured output mode can use a key-value structure, function call parameter structure, or parsable template structure. The model input package serves as the business input for the large language model to generate control instruction packages; the generated control actions, necessary repetitions, and state transition conditions are all restricted by the read-only constraint fields in the model input package.

[0070] In this embodiment, a field locking flag can also be set for read-only constraint fields. Field locking flags include three categories: non-modifiable, selectable, and completable text. Non-modifiable fields include scene identifiers, source identifiers, prohibited actions, and state transition conditions; selectable fields include allowed actions among multiple candidate sub-scenes; and completable text fields include natural language expressions in the instruction text. Field locking flags are used for field inheritance consistency verification to determine whether the control instruction package maintains consistent field sources.

[0071] Generate control instruction packages by calling a large language model based on read-only constraint fields; The large language model can be deployed locally or accessed via a controlled interface. The model input package is fed into the large language model, which generates a control instruction package based on read-only constraint fields, the current state set, system-level control rules, prohibited items, and structured output patterns. The control instruction package includes at least a scenario identifier, instruction text, control action, necessary repeating items, state transition conditions, next-round observation fields, source identifier, and confidence level. The instruction text is sent to the trainee, the control action is used to verify whether the action comes from the allowed action set, the necessary repeating items are used for repeating and comparison, the state transition conditions and next-round observation fields are used for aircraft state comparison, and the source identifier is used for training records and result verification.

[0072] The command text is generated under the constraints of permitted actions, prohibited actions, necessary repeats, and state transition conditions. For example, when the read-only constraint fields contain the command objective "go-around avoidance," the permitted action "issue go-around instruction," the necessary repeats "call sign, go-around action, target altitude," and the state transition condition "aircraft enters go-around climb state and confirms target altitude," the generated control command package contains the corresponding command text and repeats. Confidence can be formed by the model output confidence, candidate sub-scenario scores, and the completeness of structured fields. The completeness of structured fields can be determined based on the existence of required fields, whether the field type conforms to a pattern, and whether the field is parsable.

[0073] Perform field inheritance consistency checks, security boundary checks, and state reachability checks on the control instruction package; It can perform structured output pattern verification, scenario field inheritance consistency verification, numerical range verification, prohibited action verification, flight conflict verification, and state transition reachability verification. Structured output pattern verification checks whether the control instruction package contains required fields and whether the field types conform to the structured output pattern. Scenario field inheritance consistency verification checks whether the scenario identifier, source identifier, instruction target, permitted actions, prohibited actions, necessary repetitions, and state transition conditions originate from read-only constraint fields in the model input package. Numerical range verification checks whether the altitude, speed, heading, runway, frequency, and time fields are within the configured range. Prohibited action verification checks whether prohibited actions from candidate sub-scenarios or prohibited items from system-level control rules appear in the control actions and instruction text. Flight conflict verification checks whether control actions conflict with equipment states, weather conditions, conflicting targets, airspace states, or runway states in the current state set. State transition reachability verification checks whether the state transition conditions in the control instruction package can transition from the current state set to a subsequent state allowed by the hierarchical finite state machine.

[0074] When the verification fails, the error field is combined with the original model input package to form a regenerated input. If the verification passes within a limited number of times, a control instruction package is output. If the verification fails continuously, a manual takeover flag is output.

[0075] Error fields include missing fields, incorrect field types, inconsistent field inheritance, out-of-bounds values, prohibited actions, flight conflicts, or unreachable state transitions. These error fields are included as part of the regeneration input, inputting together with the original model input package into the large language model. The number of regeneration attempts can be set to 2 or 3. If the validation passes within the limited number of attempts, a control command package is output and the command text is sent; if consecutive failures occur, a manual takeover flag is output and the reason for failure is written to the state transition record. When sending the command text, the current round of model input package, control command package, validation results, and sending timestamp are simultaneously saved for repetition and comparison with the aircraft status.

[0076] In this embodiment, the automatic speech recognition model, text embedding model, and large language model can complete parameter configuration and runtime verification during the offline preparation phase. The adaptation data for the automatic speech recognition model can come from authorized civil aviation communication corpora and manually proofread text; the training objective is to minimize the sequence differences between the transcribed text and the manually proofread text. The text embedding model can use a general Chinese text embedding model or an embedding model adapted to civil aviation text; the model input is scene retrieval text or query text, and the model output is a fixed-dimensional vector. The large language model is used to generate constrained control instruction packages; the inference input is the model input package, and the inference output is a structured control instruction package. Runtime verification can use manually constructed state sets, candidate small scenarios, and model input packages to check the integrity of output fields, the inheritance of read-only constraint fields, the hit rate of prohibited actions, and the reachability of state transitions.

[0077] S105 compares the trainee's repeat slot and the aircraft's real-time parameters with the control instruction package to form an observation result package. Based on the observation result package, it controls the hierarchical finite state machine to perform state transitions, forming an updated current state set or outputting a manual takeover flag.

[0078] In this embodiment, the state closed-loop control receives control command packets, student audio or keyboard text, real-time flight parameters of the simulator after the command is issued, and the current state set. The necessary repeats, state transition conditions, and fields to be observed in the next round of the control command packet serve as the basis for repeat verification and aircraft state verification, respectively. The state closed-loop control outputs an observation result packet, an updated current state set, the next processing type, and the state transition record for this round. The next processing type can be continued retrieval, supplementary repeats, command correction, manual takeover, or termination.

[0079] Furthermore, after sending the instruction text, the student's voice or keyboard text is standardized and the call sign, altitude, heading, speed, runway, frequency, permitted action, and restriction condition slots are extracted. After the instruction text is sent, the student repetition receiving window is opened. For student voice, speech transcription, role confirmation, and text normalization are performed; for keyboard text, text normalization is performed. The normalization process includes call sign recognition, Chinese numeral conversion, unit normalization, heading formatting, frequency formatting, and action word normalization. Call sign, altitude, heading, speed, runway, frequency, permitted actions, and constraint slots are extracted from the normalized text, and for each slot, the slot value, source text, extraction confidence level, and timestamp are recorded. Slots that failed to be extracted are written to the low confidence field and used for subsequent comparisons.

[0080] The key slots are compared with the necessary repetition items for precise values ​​or configuration tolerances. The non-key expressions are compared with the instruction text in the control instruction package for semantic similarity. The real-time parameters of the aircraft are compared with the expected state transition conditions and the fields to be observed in the next round, forming an observation result package containing the repetition verification results and the aircraft state satisfaction. Key slots in necessary repeat items are compared using either exact values ​​or configuration tolerances. Call signs, runways, frequencies, and clearance actions can be compared using exact values, while altitude, heading, and speed can be compared using configuration tolerances. For example, the altitude tolerance can be set to 0, the heading tolerance to 5 degrees, and the speed tolerance to 10 knots; these values ​​are exemplary implementations and can be configured according to training rules. Non-critical expressions can be compared using keyword coverage and semantic similarity. Keyword coverage represents the ratio of the number of non-critical keywords hit in the trainee's repeat text to the number of keywords that should have been hit, and semantic similarity can be determined using the cosine similarity between the repeat text vector and the corresponding command text vector. The aircraft's real-time parameters are compared with state transition conditions and the fields to be observed in the next round to check whether the aircraft's altitude, speed, heading, vertical position, runway position, airspace position, equipment position, and conflicting targets meet the expected state transition conditions in the control instruction package.

[0081] When the recitation is completely correct and the aircraft state meets the expected state transition conditions, the hierarchical finite state machine is controlled to enter the next disposal sub-state; when the recitation is partially correct and there are no critical item conflicts, the hierarchical finite state machine is controlled to maintain the current disposal sub-state and generate a missing item prompt requirement; when the recitation is incorrect, there is a critical item conflict, or the aircraft state deviates from the expectation, the hierarchical finite state machine is controlled to generate a correction requirement or re-retrieve a requirement; when there are no qualified candidates, continuous low confidence, continuous generation failures, or the safety boundary is triggered, a manual takeover flag is output.

[0082] The repetition results are categorized as completely correct, partially correct, and incorrect. Completely correct indicates that all critical slots match the necessary repetition items, and non-critical expressions satisfy semantic constraints; partially correct indicates that there are no conflicts in critical slots, but some non-critical expressions are missing; incorrect indicates that at least one critical slot is missing, critical slot values ​​conflict, or the permitted action is incorrect. Aircraft state satisfaction is determined based on the comparison between real-time flight parameters and state transition conditions.

[0083] When the recitation is completely correct and the aircraft status meets the expected state transition conditions, the hierarchical finite state machine marks the current handling step as completed and proceeds to the next handling sub-state. If the recitation is partially correct and there are no critical item conflicts, the hierarchical finite state machine maintains the current handling sub-state and generates a missing item prompt requirement. If the recitation is incorrect or there is a critical item conflict, the hierarchical finite state machine reverts to the instruction pending generation state and generates a correction requirement; if the aircraft status deviates from the expectation but is still recoverable, the deviation field is updated and a re-retrieval requirement is generated, using the updated current state set for candidate small-scenario retrieval. If there are no qualified candidates, continuous low confidence, continuous generation failures, or safety boundary triggers, a manual takeover flag is output. When the completion conditions are met, the hierarchical finite state machine enters the special situation termination state.

[0084] Write the repetition error slots, non-critical missing expressions, and aircraft status deviation fields into the anomaly observation queue; The anomaly observation queue is used to temporarily store anomaly fields generated by comparing repetition with aircraft status. Repetition error slots include the slot name, expected value, trainee repetition value, comparison method, and error type; non-critical missing expressions include the missing item name, corresponding necessary repetition item, and source text; aircraft status deviation fields include the deviation field name, expected range, real-time value, deviation duration, and whether it is recoverable. The anomaly observation queue is sorted by the current training round identifier, status version number, and timestamp, and prioritizes reading safety boundary fields and critical slot fields.

[0085] The anomaly observation queue is read using critical slot conflicts, missing non-critical expressions, and recoverable deviations in aircraft status as anomaly categories. Based on the type of the abnormal field, the corresponding field in the abnormal observation queue is identified as a critical slot conflict, a non-critical expression missing, or a recoverable deviation in aircraft status.

[0086] The system generates correction instructions for critical slot conflicts, missing item prompts for non-critical missing expressions, and re-retrieval requests for recoverable deviations in aircraft status. The critical slot conflict category corresponds to correction instruction requirements, which include the slot name to be corrected, the expected value, the previous instruction text, and constraint fields that can send correction prompts. The non-critical missing expression category corresponds to missing item prompt requirements, which include the missing item name and the originally required repeating item. The aircraft state recoverable deviation category corresponds to re-retrieval requirements, which include the deviation field, deviation value, current state set, and safety hard constraints.

[0087] Write the generated requirement categories into the updated current state set.

[0088] The demand category is written into the updated current state set and used as input for the next round of candidate small scenario retrieval or state closed-loop control. When both critical slot conflicts and recoverable deviations in aircraft state exist simultaneously in the anomaly observation queue, after handling the critical slot conflicts, a decision is made on whether to perform a re-retrieval based on the corrected repetition and real-time flight parameters.

[0089] In this embodiment, a detectable implementation method can also be set. At the end of each round of dialogue, the training recording module records the model input packet identifier, candidate small scene identifier, control instruction packet fields, repeat slot comparison details, aircraft state satisfaction, anomaly observation queue, state version number, next processing type, and state transition record. Recording trigger conditions include control instruction packet generation, instruction text transmission, repeat result formation, aircraft state comparison completion, hierarchical finite state machine update, manual takeover flag output, and special situation termination. The records can be saved as structured logs and read via the debugging interface according to the training identifier and state version number. The structured logs are used to verify whether the control instruction packet inherits read-only constraint fields, whether the repeat comparison uses necessary repeat items, and whether the hierarchical finite state machine update meets the transition conditions.

[0090] The following description, using specific examples, elaborates on this embodiment.

[0091] In an exemplary training scenario, the offline preparation phase inputs include a runway occupancy handling data, a training review text, and a set of operating rules text.

[0092] A unified data record is generated from the data, containing the data identifier "Data A", the source location "Runway Occupation Handling Section 1", the speaking role "Controller", the standardized text "Execute Go-Around, Maintain 3000m", the data type "Training Review Text", and the version "Version 1". Candidate sub-scenarios are generated according to the handling objective "Go-Around Avoidance", the control action "Issue Go-Around Instruction", the expected repetition "Call Sign, Go-Around Action, Target Altitude", and the state transition boundary "Aircraft Enters Go-Around Climb and Confirms Target Altitude". After the rule parser confirms that the situation type, flight phase, altitude unit, and preceding steps meet the configuration conditions, the candidate sub-scenarios are determined as approved sub-scenarios and added to the situation sub-scenarios database. The sequence "Runway Occupation, Approach, Runway Unavailable, Preceding Action is Continue Approach, Handling Objective is Go-Around Avoidance" is serialized into scenario retrieval text, converted into scenario vectors by a text embedding model, and then written into the vector index.

[0093] During the online training phase, initial parameters for the training course and real-time flight parameters from the simulator are read to form a current state set. An example of a current state set is: Flight phase: Approach; Situation type: Runway occupancy; Severity: Medium; Runway status: Unavailable; Aircraft altitude: 900m; Speed: 140 knots; Heading: 180 degrees; Preceding control instruction: Continue approach; Current action objective: Go-around; Completed steps include approach clearance confirmation; Data confidence level: 0.92; State version number: 5. This current state set is then used for candidate small-scenario retrieval.

[0094] The special situation scenario library is filtered according to runway occupancy, approach, runway unavailability, and the preceding step "approach clearance confirmation." Go-around handling scenarios are retained, while scenarios requiring runway availability or landing clearance are removed. The current status set, pilot request, preceding control instructions, dialogue summary, completed steps, and current handling objective are serialized into query text and a query vector is generated. Cosine similarity is calculated between the query vector and the filtered scenario vector, and a weighted ranking is performed based on key field matching, process continuity, and data credibility. After weighted ranking, the candidate total score of the go-around handling scenario is higher than the exemplary minimum acceptance threshold of 0.65, and therefore enters the candidate scenario set.

[0095] If the prohibited action in a candidate sub-scenario corresponds to the current unavailable runway state, a hard conflict flag is generated and the corresponding candidate sub-scenario is removed. In this example, the go-around action satisfies the safety hard constraint, so the go-around handling sub-scenario is retained.

[0096] The scenario identifier, source identifier, instruction target "go-around avoidance", permitted action "issue go-around instruction", prohibited action "issue landing clearance", necessary repeating items "call sign, go-around action, target altitude", state transition condition "aircraft enters go-around climb and confirms target altitude", and completion condition "runway occupancy handling completed or track handling after go-around" from the candidate sub-scenario are written into a read-only constraint field. The read-only constraint field, together with the current state set, system-level control rules, prohibited items, and structured output mode, form the model input package.

[0097] The control instruction package is generated by calling a large language model based on the model input package. The control instruction package includes a scenario identifier, instruction text "China Southern Airlines 123, go around, maintain 3000m", control action "issue go around instruction", necessary repeat items "call sign, go around action, target altitude", state transition condition "aircraft enters go around climb and confirms target altitude", next round of observation fields "vertical status, altitude, repeat slot" and source identifier.

[0098] The control instruction packet is validated to check the completeness of structured fields, consistency between scenario and source identifiers, target altitude values ​​within the training rules' allowed range, successful verification of prohibited actions, consistency between the go-around action and runway occupancy status, and the ability of the state transition condition to move from the current handling sub-state to the go-around climb confirmation state. Once the validation is successful, the instruction text is sent.

[0099] After the trainee repeats "China Southern Airlines 123 Go-Around," the call sign and go-around maneuver slots are extracted, but the target altitude slot is missing. The call sign and go-around maneuver are compared with the necessary repeating items, resulting in a repeating result where the key slots are conflict-free and the target altitude is missing; the repeating result is partially correct. Real-time aircraft parameters show that the aircraft has entered climb mode but the target altitude has not yet been confirmed. A missing item prompt is generated, and the current handling sub-state is maintained.

[0100] After the trainee repeats "maintain 3000m", the necessary repeat items are re-compared. The call sign, go-around maneuver, and target altitude all meet the configuration conditions. The aircraft's real-time parameters show that the vertical state is climbing. The control layered finite state machine marks the go-around instruction steps as completed, updates the current state set to "go-around climb, target altitude confirmed, go-around instruction completed", and uses the updated current state set to retrieve the small scenario of track handling after the go-around.

[0101] When the subsequent steps meet the completion conditions, the hierarchical finite state machine is controlled to enter the special case termination state and the training results, review data and knowledge items to be reviewed are saved.

[0102] For the purpose of simplicity, the method steps disclosed in the above embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0103] like Figure 2 As shown, the present invention also provides a flight emergency control instruction generation and state closed-loop control system based on a large language model, comprising: The small scenario library construction module 201 is configured to extract unified data records from flight emergency data, perform processing steps segmentation and verification on the unified data records to form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index. The current state formation module 202 is configured to input the training state source, the preceding control instruction, the dialogue summary and the completed steps into the hierarchical finite state machine to form a current state set with state version numbers; The candidate small scene retrieval module 203 is configured to perform safety hard constraint filtering on the special situation small scene library based on the current state set, and perform vector similarity matching based on the vector index to output the candidate small scene set; The control instruction constraint generation and verification module 204 is configured to convert the constraint information in the candidate small scene set into read-only constraint fields, form a model input package with the current state set, call the large language model based on the model input package to generate a control instruction package, and send the instruction text after passing the safety constraint and state transition verification. The state closed-loop control module 205 is configured to compare the trainee's repeating slot and the aircraft's real-time parameters with the control instruction package to form an observation result package. Based on the observation result package, it controls a hierarchical finite state machine to perform state transitions, forming an updated current state set or outputting a manual takeover flag. It is worth noting that although only some basic functional modules are disclosed in this embodiment, it does not mean that the composition of this system is limited to the above-mentioned basic functional modules. On the contrary, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with existing technology to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. The fact that this embodiment only discloses a few basic functional modules does not mean that the scope of protection of this invention is limited to the disclosed basic functional modules. Furthermore, for ease of description, the above device is described by function as various units and modules. Of course, in implementing this invention, the functions of each unit and module can be implemented in one or more software and / or hardware.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating flight emergency control instructions and implementing state closed-loop control based on a large language model, characterized in that, include: Extract unified data records from flight emergency data, segment and verify the unified data records for processing steps to form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index; The training state source, preceding control instructions, dialogue summary, and completed steps are input into the hierarchical finite state machine to form the current state set with state version numbers; Perform hard-constraint filtering on the special scenario library based on the current state set, and perform vector similarity matching based on vector index to output a candidate scenario set; The constraint information in the candidate small scene set is converted into read-only constraint fields, which are combined with the current state set to form a model input package. Based on the model input package, the large language model is called to generate a control instruction package. After passing the safety constraint and state transition verification, the instruction text is sent. The trainee repeats the slot and the aircraft's real-time parameters, respectively, and compares them with the control instruction package to form an observation result package. Based on the observation result package, the hierarchical finite state machine is controlled to perform state transitions, forming an updated current state set or outputting a manual takeover flag.

2. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, Extracting standardized data records from flight incident data further includes: The system performs noise reduction, voice activity detection, speaker separation, or vocal tract annotation on the call audio, and uses an automatic speech recognition model adapted to the civil aviation communication corpus to generate transcribed text with start and end times and confidence scores. Match call sign, frequency, turn order, and speaker characteristics with the transcribed text to label controller or pilot roles. Call signs, Chinese numerals, units, and civil aviation terminology are standardized, and low-confidence transcribed texts are sent to a manual sampling queue. The output is a unified data record containing data identifier, source location, speaker role, start and end time, standardized text, confidence level, data type, and version.

3. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, The unified data record processing steps are segmented and verified to form approved sub-scenarios, which further include: The conversation turns are segmented based on speaker changes, silence intervals, and timestamps. The special situation type, applicable flight phase, pilot request, instruction target, permitted actions, prohibited actions, necessary repeats, state transition conditions, and completion conditions are extracted from the conversation turns, report text, debriefing text, and operation rule text. The extracted fields within the same processing step are combined into candidate sub-scenes. The enumeration values, units, numerical ranges, dependencies, and source integrity are verified. After verification and review, the approved sub-scenes are output.

4. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, The structured constraint information of the approved small scenarios is written into the special scenario library, and the corresponding scenario vector is written into the vector index, further including: Store state constraints, necessary repetitions, and state transition conditions in a key-value structure; store source extracts and example expressions in text format; and connect adjacent processing steps using predecessor and successor identifiers. The special situation type, applicable flight phase, state constraints, pilot request, preceding actions and handling objectives are serialized into scene retrieval text, and then input into the text embedding model to obtain scene vectors. Write the scene vector along with the sub-scene identifier, document version, review version, and embedded model version into the vector index.

5. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, The training state source, preceding control instructions, dialogue summary, and completed steps are input into a hierarchical finite state machine to form a current state set with state version numbers, which further includes: Obtain the initial parameters of the training course or the previous round of updated status, the simulator's real-time flight parameters, pilot requests, preceding control instructions, summaries of the most recent rounds of dialogue, and completed steps, and input them into a hierarchical finite state machine; Constrained by the transition conditions of the top-level emergency process status and emergency handling sub-states, the flight phase, emergency type, severity, airspace or runway, flight altitude, flight speed, flight heading, equipment status, weather conditions, conflict target, current handling target, remaining handling time, and data confidence are written into the current state set.

6. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, The current state set is used to perform safety hard constraint filtering on the special situation mini-scenario library, and vector similarity matching is performed based on vector indexes to output a candidate mini-scenario set, which further includes: Using special situation type, flight phase, equipment availability, prohibited actions, safety boundaries and preceding steps as hard constraints, retain small scenarios that meet the conditions of enumeration inclusion, continuous quantity tolerance, timing sequence and duration, and eliminate small scenarios that do not meet the conditions of enumeration inclusion, continuous quantity tolerance, timing sequence and duration. The current status set, pilot request, preceding control instructions, dialogue summary, completed steps and current action target are serialized into query text and converted into query vectors. The cosine similarity between the query vector and the filtered scene vector is calculated, and the cosine similarity is used as the semantic similarity. The filtered small scenes are then weighted and sorted based on semantic similarity, key field matching degree, process continuity and data credibility, and a preset number of candidate small scenes with a value higher than the minimum acceptance threshold are output. If no qualified candidates are found, expand the scope of non-critical searches while retaining hard safety constraints. If no candidates are found, output a manual takeover flag.

7. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, The constraint information in the candidate small scene set is converted into read-only constraint fields, which are then combined with the current state set to form a model input packet. Based on the model input packet, a large language model is called to generate a control instruction packet. After passing safety constraint and state transition verification, the instruction text is sent, which further includes: Select at least one candidate scenario from the candidate scenario set that has the same source version and whose state constraints do not conflict. Write the scenario identifier, source identifier, instruction target, allowed action, prohibited action, necessary repetition, state transition condition and completion condition into the read-only constraint field. The model input package is composed of read-only constraint fields, current state set, system-level control rules, prohibited items, and structured output patterns, and is used as input for the large language model to generate control instruction packages; The large language model is called based on read-only constraint fields to generate control instruction packages. Field inheritance consistency checks, security boundary checks, and state reachability checks are performed on the control instruction packages. After the checks pass, the instruction text is sent.

8. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 1, characterized in that, The trainee repeats the slot and the aircraft's real-time parameters, respectively, and compares them with the control instruction packet to form an observation result packet. Based on the observation result packet, the hierarchical finite state machine is controlled to perform state transitions, forming an updated current state set or outputting a manual takeover flag, which further includes: After sending the instruction text, the student's voice or keyboard text is standardized and the call sign, altitude, heading, speed, runway, frequency, permitted action and restriction condition slots are extracted; The key slots are compared with the necessary repetition items for precise values ​​or configuration tolerances. The non-key expressions are compared with the instruction text in the control instruction package for semantic similarity. The real-time parameters of the aircraft are compared with the expected state transition conditions and the fields to be observed in the next round, forming an observation result package containing the repetition verification results and the aircraft state satisfaction. When the repetition is completely correct and the aircraft state meets the expected state transition conditions, the hierarchical finite state machine controls the next disposal sub-state. When the recitation is correct and there are no key item conflicts, the hierarchical finite state machine is controlled to maintain the current processing sub-state and generate missing item prompts. When there is a repetition error, a critical item conflict, or the aircraft state deviates from the expected state, the hierarchical finite state machine controls the generation of corrective requirements or the re-retrieval of requirements. Output a manual takeover flag when there are no qualified candidates, continuous low confidence, continuous generation failures, or when a security boundary is triggered.

9. The method for generating flight emergency control instructions and implementing closed-loop state control based on a large language model according to claim 8, characterized in that, When the recitation is correct and there are no key item conflicts, the hierarchical finite state machine is controlled to maintain the current processing sub-state and generate missing item prompts. When there is a repetition error, a critical item conflict, or the aircraft state deviates from expectations, the hierarchical finite state machine generates corrective requirements or retrieves new requirements, further including: Write the repetition error slots, non-critical missing expressions, and aircraft status deviation fields into the anomaly observation queue; The anomaly observation queue is read using critical slot conflicts, missing non-critical expressions, and recoverable deviations in aircraft status as anomaly categories. The system generates correction instructions for critical slot conflicts, missing item prompts for non-critical missing expressions, and re-retrieval requests for recoverable deviations in aircraft status. Write the generated requirement categories into the updated current state set.

10. A flight emergency control command generation and state closed-loop control system based on a large language model, characterized in that, include: The small scenario library construction module is configured to extract unified data records from flight emergency data, perform processing steps and verification on the unified data records to form approved small scenarios, write the structured constraint information of the approved small scenarios into the emergency small scenario library, and write the corresponding scenario vector into the vector index. The current state formation module is configured to input the training state source, preceding control instructions, dialogue summary and completed steps into a hierarchical finite state machine to form a current state set with state version numbers; The candidate small scene retrieval module is configured to perform safety hard constraint filtering on the special situation small scene library based on the current state set, and perform vector similarity matching based on the vector index to output the candidate small scene set; The control instruction constraint generation and verification module is configured to convert the constraint information in the candidate small scene set into read-only constraint fields, form a model input package with the current state set, call the large language model based on the model input package to generate a control instruction package, and send the instruction text after passing the safety constraint and state transition verification. The state closed-loop control module is configured to compare the trainee's repeat slot and the aircraft's real-time parameters with the control instruction package to form an observation result package. Based on the observation result package, it controls the hierarchical finite state machine to perform state transitions, forming an updated current state set or outputting a manual takeover flag.

Images