A method for automatically generating special scenarios during flight training

Abstract

This invention discloses an automatic generation method for flight training emergency scenarios, relating to the field of flight training technology. It includes the following steps: Step 1: Constructing an integrated flight simulation environment for training and evaluation, and establishing a unified and standardized emergency injection interface to obtain the basic carrier for emergency scenario simulation operation and injection; Step 2: Establishing a multi-dimensional parameterized description model and hierarchical parameter space for emergency scenarios, structurally representing individual emergency elements as a unified four-tuple containing the target object, emergency type, injection method, and timing parameters, thus obtaining a standardized description specification for emergency scenarios. This invention solves the problems of traditional emergency scenario design—high reliance on manual labor, high manpower input, low output efficiency, and limited scenario types—by modeling emergency scenario generation as a multi-step sequential decision-making process and combining reinforcement learning to build an intelligent agent for emergency scenario generation and a closed-loop automatic generation mechanism.

Description

Technical Field

[0001] This invention relates to the field of flight training technology, specifically to a method for automatically generating special flight training scenarios. Background Technology

[0002] With the widespread application of flight simulators, flight trainers, and various simulation training systems, flight training is gradually evolving from simple flight skills training to training in emergency response capabilities, comprehensive judgment capabilities, and resource management capabilities.

[0003] During flight training, emergency scenarios (such as system failures and environmental anomalies) are an important means of testing pilots' emergency response capabilities. Existing flight training systems typically support the injection of various preset emergency scenarios into the simulation environment for pilot training and assessment.

[0004] However, as the complexity of flight systems continues to increase, flight training places higher demands on the diversity, complexity, and relevance of emergency scenarios. The traditional approach of relying on human experience to design emergency scenarios is gradually becoming insufficient to meet training needs. Therefore, this paper proposes an automatic generation method for flight training emergency scenarios to address this problem. Summary of the Invention

[0005] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for automatically generating special flight training scenarios, thus solving the problems mentioned in the background section.

[0006] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a method for automatically generating special flight training scenarios, comprising the following steps: Step 1: Construct an integrated flight simulation operation environment for training and evaluation, and build a unified and standardized special situation injection interface to obtain the basic carrier for special situation scenario simulation operation and injection. Step 2: Establish a multi-dimensional parameterized description model and hierarchical parameter space for special situation scenarios. Represent a single special situation element in a structured way as a unified quadruple containing the target object, special situation type, injection method, and timing parameters, thereby obtaining a standardized description specification for special situation scenarios. Step 3: Model the special situation scenario generation process as a multi-step sequential decision-making process to obtain a standardized generation framework for the special situation scenario sequence; Step 4: Construct a multi-dimensional comprehensive evaluation index for training effectiveness, and based on the evaluation index, form a comprehensive reward function to quantify the training value of special scenarios, thereby obtaining the quantitative standard for the training value of special scenarios and the optimization goal of reinforcement learning. Step 5: Construct an intelligent agent for generating special situation scenarios based on reinforcement learning. With the comprehensive reward function as the optimization objective, and combining the simulation environment, the standardized description specification of special situation scenarios, and the sequence generation framework, construct a closed-loop automatic generation mechanism for special situation scenarios and output a batch of candidate special situation scenarios. Step 6: Conduct quantitative evaluation of the training value of candidate special scenarios, similarity analysis and redundancy elimination, and screen out special scenarios with high training value.

[0007] Preferably, the construction steps of the basic carrier for the simulation and injection of the special situation scenario are as follows: The integrated flight simulation environment integrates a flight dynamics model adapted to all flight phases, an aircraft system model with an injectable interface, an environment model that supports time-series evolution, a multi-mode human-machine interaction interface, and a full-process data acquisition and recording module. The established special situation injection interface supports standardized structured input of special situations, injection of multiple types of special situation evolution modes, multi-condition combination triggering and recovery control, injection conflict and mutual exclusion rule verification, and also has the ability to trace the entire special situation injection process and reproduce scenarios.

[0008] Preferably, the steps for constructing the standardized description specification for the special situation scenario are as follows: A multidimensional parameterized description model is constructed using a hierarchical modeling approach, which sequentially sets up an object layer, a type layer, an injection method layer, and a timing layer. The object layer limits the scope of the special situation, the type layer defines the anomaly category of the special situation, the injection method layer clarifies the implementation model of the special situation in the simulation environment, and the timing layer defines the triggering, evolution, and recovery rules of the special situation. Finally, a single situation element is structured as a unified four-tuple structure containing the target, situation type, injection method, and timing parameters.

[0009] Preferably, the construction steps of the hierarchical parameter space are as follows: Set up three types of parameter sets: system special situation parameter set, environmental special situation parameter set, and timing parameter set; The system incident parameter set is used to describe the abnormal behavior and performance degradation characteristics of various aircraft systems, the environmental incident parameter set is used to describe the abnormal and disturbance characteristics of external flight conditions, and the time series parameter set is used to describe the triggering, duration, evolution, and recovery rules of incidents in the time dimension.

[0010] Preferably, the execution steps of the standardized generation framework for the special situation scene sequence are as follows: The complete sequence of special situation scenarios is represented as a sequence of multiple special situation elements in chronological order, and a single-step decision corresponds to the generation of a single special situation element; The single-step decision-making process sequentially executes four stages: state reading, candidate generation, constraint filtering, and selection and recording. State reading is used to obtain the current flight status, system mode, and incident history. Candidate generation is used to generate candidate incident elements from the hierarchical parameter space. Constraint filtering is used to eliminate candidates that do not meet physical rules and training requirements. Selection and recording is used to determine the final injected incident elements and complete the structured information retention.

[0011] Preferably, the quantitative standard for the training value of the special situation scenario and the steps for constructing the reinforcement learning optimization objective are as follows: Four core evaluation indicators are set: flight quality deviation indicators, safety margin indicators, control requirements and load indicators, and training procedure compliance indicators. Among them, the flight quality deviation index is used to measure the cumulative error and persistence of the aircraft's state deviating from the target state; the safety margin index is used to measure the degree to which the flight process approaches or exceeds the safety boundary; the control requirements and load index is used to measure the control intensity required to complete the flight mission; and the training procedure compliance index is used to measure the degree to which the scenario triggers the key points of the training outline. A comprehensive reward function with out-of-bounds penalty terms is constructed based on four core indicators, which serves as the optimization objective for reinforcement learning.

[0012] Preferably, the construction steps of the special situation generation agent are as follows: The input to the agent is a concatenated state vector, which contains four types of information: flight state, control input, training phase, special situation history, and difficulty estimation. The agent's output is a mixed discrete and continuous action vector, where the discrete part is used to determine the target, type, injection method, and triggering method of the special situation, and the continuous part is used to determine the intensity parameters and timing parameters of the special situation; The agent continuously optimizes the special situation generation strategy through reinforcement learning algorithms, guided by the comprehensive reward function.

[0013] Preferably, the single-round execution steps of the special situation closed-loop automatic generation mechanism are as follows: The special situation scenario generation agent outputs a special situation scenario sequence according to the current strategy, writes it into a standardized scenario injection plan file, and generates random seeds to achieve scenario reproducibility. The simulation environment executes flight missions according to preset switchable control reference modes, and completes emergency injection and full-process simulation evolution according to the injection plan; it collects and records all flight status, control inputs, system modes, alarm signals, and emergency injection data throughout the simulation process. After the simulation, the evaluation index is calculated and the comprehensive reward value is settled. Invalid scenarios that are out of bounds, untrainable, or have abnormal values ​​are marked simultaneously. Based on the valid samples of this round, the agent's strategy is optimized, and batch special scenarios are automatically generated through repeated iterations.

[0014] Preferably, the quantitative evaluation of the training value of the candidate special situation scenarios is performed using the following steps: Based on multi-dimensional comprehensive evaluation indicators, the simulation operation data of each candidate special situation scenario is normalized, the analysis window is divided according to the flight stage, and the comprehensive reward value of a single scenario is calculated by combining the statistical results of key events. The training value is then quantitatively ranked, and low-value scenarios with comprehensive reward values ​​lower than the preset threshold are eliminated.

[0015] Preferably, the steps for similarity analysis, redundancy removal, and difficulty grading of the candidate special situation scenarios are as follows: Similarity analysis is conducted from two dimensions: generated parameters and simulation results. Parameter similarity is based on the structural information of the special situation sequence to construct a weighted distance metric to complete coarse clustering, while result similarity is based on the time series data of the simulation output to complete fine screening through statistical feature matching. Based on the similarity analysis results, redundant scenarios are eliminated and representative samples with the highest training value among similar scenarios are retained. Based on the structural complexity, safety margin, and operational load of the scenarios, the selected special situations are divided into three progressive difficulty levels: basic, intermediate, and advanced.

[0016] Beneficial effects The present invention has the following beneficial effects: (1) This method for automatically generating special flight training scenarios solves the problems of traditional special scenario design being highly dependent on manual labor, having a large amount of manpower, low output efficiency, and a single scenario type by modeling the generation of special scenarios as a multi-step sequential decision-making process and combining reinforcement learning to build a special scenario generation agent and a closed-loop automatic generation mechanism. It achieves the automatic generation of diverse flight training special scenarios without the need for a large number of manual interventions.

[0017] (2) This method for automatically generating special flight training scenarios solves the problems of traditional manual scenario design lacking unified evaluation standards, training value being unquantifiable, and horizontal comparison between different scenarios by constructing a multi-dimensional comprehensive evaluation index system oriented towards training effect. It achieves accurate quantification and comparison of the training value of special scenarios.

[0018] (3) This method for automatically generating flight training emergency scenarios solves the problems of repetitive redundancy, inability to systematically retain high-value scenarios, and poor reusability of traditional manually designed scenarios by setting up a dual-dimensional similarity analysis, a redundancy elimination mechanism, and a standardized scenario library management system. It realizes the automatic screening and retention of high-training-value emergency scenarios and forms a reusable emergency scenario library.

[0019] (4) This method for automatically generating special scenarios for flight training solves the problems of incomplete training coverage, weak scenario targeting, poor consistency of training in different batches, and high scenario design and maintenance costs under the traditional manual design mode by automatically generating special scenarios throughout the entire process, standardizing management and control and hierarchical screening system. It improves the coverage, targeting and consistency of flight training and reduces the design and maintenance costs of special scenarios.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for automatically generating special flight training scenarios according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention provides a technical solution: a method for automatically generating special flight training scenarios, such as... Figure 1 As shown, the specific steps are as follows: Step 1: Construct an integrated flight simulation operation environment for training and evaluation, and build a unified and standardized special situation injection interface to obtain the basic carrier for special situation scenario simulation operation and injection.

[0024] A simulation environment for flight training and emergency assessment is constructed. This simulation environment serves as the operational foundation for all subsequent processes, including emergency injection, scenario generation, and evaluation analysis. Its design goal extends beyond simply meeting routine flight training requirements; it aims to operate stably under various complex operating conditions, system states, and environmental conditions, realistically reflecting the dynamic response characteristics of the aircraft and its systems. In this embodiment, the simulation environment is built upon a six-degree-of-freedom nonlinear flight dynamics model that conforms to the civil aviation high-level flight simulator certification standards. It is compatible with mainstream civil aviation aircraft such as the A320, B737, C909, and C919. The simulation step size is set to 10ms, meeting the relevant requirements of Civil Aviation Administration of China (CCAR-60) for high-level training equipment.

[0025] The simulation environment integrates a flight dynamics model adapted to all flight phases, an aircraft system model with an injectable interface, an environment model that supports time-series evolution, a multi-mode human-machine interface, and a full-process data acquisition and recording module.

[0026] The flight dynamics model, acting as the state generator of the simulation environment, is responsible for calculating aircraft state changes within each simulation step and outputting state variables such as attitude, velocity, trajectory, position, altitude, and acceleration. The output of the dynamics model must meet the common requirements of the system model, environment model, and evaluation module. Attitude data is not only used for display but can also be referenced by flight control laws, alarm logic, and stability criteria. Velocity data is used for flight evaluation and can also be used for calculating certain special situation triggering conditions. Flight phases must cover the basic motion characteristics of takeoff, climb, cruise, maneuvering, approach, go-around, landing, and taxiing, and support continuity during phase transitions.

[0027] Aircraft system models are used to describe the functions, logic, and dynamic characteristics of systems such as the power plant, flight control, hydraulics, power supply, and avionics. To support subsequent situation injection and sequence generation, the system model needs to have injectable points and parameterizable interfaces. The power plant model can provide simulation capabilities for thrust / power output, response hysteresis, and limiting logic to represent situations such as response delays and sensor anomalies. The flight control system model can provide characteristics such as control law switching, actuator jamming / speed limiting / failure, and can output key state quantities such as control commands and control surface positions. The hydraulic system model can provide characteristics such as multi-channel hydraulic source switching and engine-driven pump or converter failure, affecting multiple actuators in the same channel. The power supply system model can provide characteristics such as primary / backup power switching and load distribution, and can provide simulation effects such as busbar or converter failure, affecting cascaded equipment failure or degraded operation. The avionics system model implements the logic of the actual aircraft model and has injectable alarm triggering logic, mode switching logic, and state output capabilities, providing data sources for training program compliance evaluation.

[0028] The environmental model provides inputs for external conditions, including atmosphere, wind, icing, turbulence, and pavement conditions. The environmental model must support the spatial and temporal evolution of special situation parameters. Environmental variables include both stable, configurable baseline values ​​and disturbances that vary over time or space. For wind fields, random disturbances such as gusts and turbulence can be superimposed on a baseline with constant values ​​or altitude profiles. For pavement pavements, with a piecewise constant friction coefficient as the baseline, local or temporal variations caused by rainfall, icing, etc., can be superimposed. The environmental system persists throughout the entire mission phase; therefore, the environmental model should provide a unified output interface and variable definitions to avoid splitting into multiple different models for each phase, ensuring consistent access and continuous evolution throughout the entire mission.

[0029] The human-computer interface provides input sources, including joysticks / wheels, pedals, accelerator, and brakes, and also supports automatic operation or playback of input. To meet the needs of batch generation and comparability analysis, the interface supports three modes: manual operation mode for human-assisted training and verification, automatic operation / automatic control mode for batch operation and stable comparison, and operation data playback mode for controlled experiments with the same input but different special conditions, significantly improving the comparability between scenarios.

[0030] The data acquisition and recording module is the foundation for evaluating metric calculations, similarity analysis, screening and grading, and data storage and reproduction. Besides recording state quantities and manipulation inputs, it must include three core types of information: key event timestamps, injected parameter snapshots, and quality markers. Key event timestamps include mode switching, alarm occurrence / clearance, special situation triggering / recovery, and stage switching. Injected parameter snapshots include parameter values ​​at the time of special situation triggering and key node values ​​during the evolution process. Quality markers include whether the simulation is numerically stable, whether model errors have occurred, and whether the untrainable criterion has been triggered. The recorded data logs will be stored in a structured format for subsequent automated processing and retrieval.

[0031] Building a unified incident injection interface within the simulation environment is crucial for decoupling the automatic generation of incident scenarios from simulation execution. This interface should not only support the injection of specific faults into the system but also comprehensively describe the fault's introduction method and triggering conditions, its dynamic evolution over time, recovery mechanisms and applicable conditions, and the consistent reproducibility of the scenario across different playbacks. Incident elements are described using a standardized structure, containing information such as the target object, incident type, injection method, and timing parameters, to fully describe the content and evolution of a single incident. The incident injection interface is responsible for performing anomaly injection in the corresponding system or environment model based on the information from the incident elements and controlling its triggering, duration, and recovery processes.

[0032] For common emergency evolution patterns, templates are used to abstract and support reuse and expansion, including templates for step, ramp, hysteresis, stochastic processes, and state machines. The step template directly switches parameters to the abnormal state at the trigger moment; the ramp template changes linearly or piecewise linearly over a set period to simulate gradual deterioration; the hysteresis template characterizes system inertia and response delay; the stochastic process template simulates noise enhancement, random jumps, or intermittent failures; and the state machine template simulates mode switching, repeated triggering of protection logic, or intermittent interruptions.

[0033] The triggering logic supports multiple combinations of conditions, not just fixed-time triggers, including stage triggers, threshold triggers, and system status triggers. Stage triggers include those triggered after takeoff, after mode activation, or after a critical event. Threshold triggers occur when altitude, speed, attitude, or deviation exceeds a set range. System status triggers include those triggered upon entering a specific control law or several seconds after an alarm is triggered. The recovery mechanism supports automatic recovery, manual deactivation, or condition-based automatic recovery.

[0034] In scenarios with multiple special cases overlapping, logical conflicts or combinations that do not conform to physical laws may occur. To address this, three types of rules are introduced at the interface layer: priority, mutual exclusion constraints, and consistency checks. The priority rule determines whether to merge or cover cases affected by multiple injections based on a strategy. The mutual exclusion constraint rule restricts the simultaneous existence of special cases or requires a time interval. The consistency check rule verifies the parameter range and system state before injection to avoid numerical divergence or abnormal behavior.

[0035] To ensure the reproducibility of the scenario, each special injection should record the injection item ID, the target object, the injection type and complete parameter set, the trigger and recovery timestamps and corresponding conditions, key evolution nodes, and the injection execution status. These records should form a complete structured log to support replay reproduction and interpretable analysis.

[0036] Step 2: Establish a multi-dimensional parameterized description model and hierarchical parameter space for special situation scenarios. Represent a single special situation element in a structured way as a unified quadruple containing the target object, special situation type, injection method, and timing parameters, thereby obtaining a standardized description specification for special situation scenarios.

[0037] A complete parameter space for special scenarios is established to provide a structured description of system anomalies, environmental anomalies, and their occurrence processes that may occur during training. Once special scenarios are formalized as elements in the parameter space, whether manually orchestrated, rule-generated, or automatically generated by reinforcement learning, they can all be performed under the same descriptive system, thereby achieving scalable generation and comparable evaluation.

[0038] The parameter space for special situations consists of three sets of parameters: system special situation parameters, environmental special situation parameters, and time-series parameters, constructed using a hierarchical modeling approach. Level A is the object layer, defining the objects affected by the special situation, such as the power plant, flight control system, hydraulic system, avionics system, pavement model, and wind field model. The object layer limits the scope of the special situation, preventing the same fault from appearing in different system modules under different variable names. Level B is the type layer, defining the types of special situations, such as failure, degradation, bias, drift, delay, increased noise, efficiency decay, and jamming. The type layer ensures that special situations have a unified semantics, allowing the same type of abnormal fault to be reused across multiple systems or components. Level C is the injection layer, defining the injection method and model parameters, i.e., how the special situation is represented in the simulation. Typical methods include parameter scaling, parameter bias, dynamic hysteresis, first-order / second-order models, stochastic process-driven simulation, logic gating, and state machine forced switching. Model parameters include intensity coefficients, bias amplitude, time constants, and upper and lower bound constraints. Level D is the timing layer, which defines the timing parameters, including when the incident is triggered, its duration, its evolution process, and its recovery method. The timing layer transforms static anomalies into dynamic processes, enabling the scenario to reflect real operational characteristics such as gradual changes, intermittent occurrences, or phased recovery.

[0039] Therefore, a special situation element can be uniformly represented as a quadruple: In the formula: The standardized, structured representation of a single emergency element is the smallest unit that constitutes a complete emergency scenario; Indicates the target of the special situation, used to clarify the target module, subsystem, component or external environment process into which the special situation is injected; Indicates the incident type, used to define the anomaly or degradation category corresponding to the incident; The method of injecting special situations is used to clarify the numerical implementation model and evolution mode of special situations in the simulation environment; The timing parameters represent special situations, used to characterize the triggering, duration, evolution, and recovery rules of special situations over time.

[0040] System Special Parameter Set Used to describe abnormal behavior or performance degradation of aircraft systems, and to meet the requirements of automatic generation and fine control, these situations will be expanded from discrete quantities to a set of mixed parameters with adjustable intensity, rate, threshold, and duration. Typical categories include power plant, flight control, hydraulic / power supply, and avionics systems. Power plant situations describe anomalies related to the engine and its control system, such as thrust response delay, maximum thrust ratio reduction, climb thrust limit changes, and sensor link anomalies. Flight control situations describe anomalies of the flight control system and actuators, such as actuator efficiency degradation, reduced maximum control surface deflection ratio, control surface rate limitation, and flight control law degradation. Hydraulics and power supply situations describe anomalies of the airborne energy system, such as single-channel hydraulic pressure drop, power system load unloading, and intermittent system function loss due to power failure. Avionics system situations describe logical or informational anomalies in the avionics system, such as alarm trigger delay, alarm trigger threshold drift, mode switching anomalies, and mode jitter.

[0041] Environmental Special Parameter Set Used to describe abnormal or increased challenges in external conditions, environmental situations differ from system situations in that they are characterized by randomness and time-varying nature. They are categorized into four main types: wind field and turbulence, meteorology and atmosphere, runway surface, and visual conditions. Wind field and turbulence includes mean wind speed and direction, gradient wind profiles varying with altitude, gust amplitude, duration, and probability of occurrence, turbulence intensity level, spectral parameters, and spatial correlation length, wind shear intensity and rate of change of direction, wind shear trigger height window, and duration range. Meteorology and atmosphere include temperature deviation, pressure-altitude deviation, icing characteristics, icing intensity, icing growth rate, and icing de-icing conditions. Runway surface includes runway contaminant status, runway friction conditions, runway slope, and terrain elevation. Visual conditions include visibility, cloud base, and illumination.

[0042] Timing parameter set This section describes the occurrence time, evolution process, and end time of an emergency, including triggering conditions, duration, evolution process, recovery conditions, and constraint rules. Triggering conditions can be described by combinations of time, altitude, distance, flight stage, mode state, and event signals, such as a certain time after entering a stage, altitude falling below a threshold, activation of a specific mode, or reaching a certain alarm. Duration can be fixed, random, or continue until the recovery condition is met. Evolution process includes step, ramp, and piecewise linear events. Recovery conditions include automatic recovery, manual deactivation, or condition-triggered recovery. Constraint rules include rules specifying that certain emergencies must occur first, cannot occur simultaneously, or must occur at certain intervals.

[0043] Step 3: Model the special situation scenario generation process as a multi-step sequential decision-making process to obtain a standardized generation framework for the special situation scenario sequence.

[0044] The process of generating special scenarios is modeled as a multi-step decision-making process, with each step outputting a special scenario element, thus forming a complete sequence of special scenarios.

[0045] Special situation sequence is used to describe the process of multiple special situations occurring sequentially or superimposed in time during a complete simulation. A special situation is represented as a sequence of K special situation elements.

[0046] In this embodiment, the maximum number of special situation injection steps K in a single training session is preset to 4 to avoid excessive accumulation of special situations that could lead to uncontrollable simulation.

[0047] Two types of boundaries are defined: sequence length boundaries and trainable boundaries. Sequence length boundaries are used to prevent the superposition of too many special cases from causing uncontrollable simulations, difficulties in training convergence, or excessive computational costs. Trainable boundaries are used to filter theoretically existing combinations that are not meaningful for training or are obviously unmanageable, and to exclude them in advance through constraint mechanisms. For example, some system degradations are only allowed to occur after a specific stage, and if the superposition of a certain environmental special case and a certain system special case will lead to obvious instability of the system, then mutual exclusion constraints should be set to eliminate them.

[0048] The special situation generation process can be regarded as a multi-step decision-making process. Each step determines whether to inject a new special situation and how to inject it based on the current simulation state. The whole process executes four stages in sequence: state reading, candidate generation, constraint filtering, selection and recording.

[0049] The status reading phase is used to obtain the current flight status, system mode, and incident history. The specific content read includes: current flight status, such as altitude, airspeed, sideslip angle, attitude and its rate of change; system mode status, such as autopilot / flight control mode, thrust / autothrottle mode, and health status of critical systems; incident history and current incident list, such as the objects of incidents, the type of each incident, the current evolution stage, the current cumulative number of incidents injected, and the estimated difficulty level.

[0050] The candidate generation stage generates candidate situation elements from the hierarchical parameter space. This can be achieved in various ways, including rule-based enumeration, random sampling of the parameter space, and reinforcement learning policy output. Rule-based enumeration limits the types and objects of consideration based on the current flight phase and mode; for example, prioritizing wind shear and autothrottle anomalies during the approach and landing phase. Random sampling of the parameter space randomly samples parameters such as intensity, trigger altitude, and duration within the limited set of objects or types, forming several candidate versions; for example, sampling intensity, range altitude, and duration for wind shear. Reinforcement learning policy output involves the agent directly outputting complete parameters.

[0051] The constraint filtering stage is used to eliminate candidates that do not conform to physical rules and training requirements. It performs multi-level filtering on the candidate set, which can be broken down into physical rationality checks, training rationality verification, and mutual exclusion and conflict elimination. For example, it checks whether parameters are within the physically reachable range, whether the rudder deflection angle does not exceed the limit, and whether the thrust is not negative; whether it is obviously unmanageable, such as superimposed extreme situations under extreme boundary conditions; and directly eliminates situations that are mutually exclusive with already injected situations. The same system will not experience both complete failure and normal degradation.

[0052] The selection and recording stage is used to determine the special situation elements to be injected in the final stage and to complete the structured information retention. The special situation elements to be injected in this step are selected from the set after constraint filtering. The selected special situation elements need to be recorded, including scene ID, element number, object, type, injection method, timing parameters, etc.

[0053] When using reinforcement learning to generate special situations, the scenario generation can be regarded as a sequential decision problem. The agent needs to output the action of how to inject the special situation next based on the current scenario state. For this purpose, the state vector, action vector and termination condition need to be clearly defined.

[0054] The state space design includes four types of sub-vectors: flight state sub-vector, control input and change sub-vector, training phase and time sub-vector, and situation history and difficulty estimation sub-vector.

[0055] Flight state subvector , including height ,airspeed Angle of attack Sideslip angle Roll angle Pitch angle Heading angle .

[0056] Manipulating inputs and changing subvectors Includes elevator control Aileron control rudder control Thrust control And the corresponding rate of change of manipulation.

[0057] Training phase and time subvectors Includes flight training phase markings This indicates that the current flight phase is any of the following: takeoff, climb, cruise, descent, approach, or go-around, as well as the time within each phase. .

[0058] Special situation history and difficulty estimation subvector Including the number of injected special cases Current Special Situation Code Scene difficulty level .

[0059] All the above information is concatenated into a high-dimensional vector, which serves as the input to the agent at the decision step.

[0060] The action space design employs a hybrid of discrete and continuous actions, fully specifying the next situation element. The discrete component includes object selection (choosing one from a given set of objects such as aircraft systems or environmental parameters); type selection (fault, degradation, bias, drift, delay, etc.); and injection method (step, ramp, first-order dynamic, stochastic process, etc.). The continuous component includes intensity parameters (bias, dead time, time constant, drift rate, etc.); and timing parameters (trigger time, trigger altitude, duration, change slope, etc.).

[0061] To ensure training stability and the security of generated results, clear termination and protection conditions are defined in the decision-making process, including a decision step limit, safety boundary termination, numerical stability protection, and training effectiveness protection. The decision step limit is set at the maximum number of scenarios; after this, no new scenarios are injected, and the simulation continues until completion. Safety boundary termination occurs if the simulation significantly exceeds the safety envelope, such as overload, excessive attitude, or failure to recover after stall; the round can be terminated immediately with a strong penalty. Numerical stability protection considers numerical divergence or non-physical values ​​as unusable samples and marks them as invalid rounds. Training effectiveness protection allows for early termination of a round if there are no significant deviations or critical events triggered over a relatively long period, and the training objective is challenging enough, with a light penalty or no reward.

[0062] To decouple the scenario generation logic from the simulation engine, a unified structured file is used to describe each scenario. A complete scenario injection file includes basic scenario information, initial conditions and environment, a scenario sequence list, constraints, and verification results. Basic scenario information includes a unique scenario identifier and version number. Initial conditions and environment include weight, center of gravity position, configuration, flaps, landing gear status, initial altitude, speed, heading and position, and atmospheric and wind field parameters. Each scenario element in the scenario sequence list includes an index number, an object (target), a scenario type, an injection method, specific intensity and model parameters (params), a schedule trigger condition, an evolutionary model description, recovery conditions, constraints_passed mutual exclusion rationality verification results, and optional notes. Constraints and verification results include whether the global mutual exclusion rule check was passed and whether the constraints meet the specified difficulty level.

[0063] The emergency sequence list consists of multiple emergency elements arranged in chronological order. Each emergency element can be defined as: {"index":i, "object":o, "type":t, "method":m, "params":{...}, "schedule":{ "trigger":"expression or structured condition", "profile":"Evolutionary model description", "stop": "Restore conditional expression" }, "constraints_passed":true / false, "notes": "Optional notes" } Step 4: Construct a multi-dimensional comprehensive evaluation index for training effectiveness, and based on the evaluation index, form a comprehensive reward function to quantify the training value of special scenarios, thereby obtaining the quantitative standard of the training value of special scenarios and the optimization goal of reinforcement learning.

[0064] Under the automatic generation mechanism, the system will generate a large number of candidate scenarios. However, not all scenarios have training value. Therefore, it is necessary to build an automatic calculation and interpretable comparison index system to comprehensively evaluate the performance of scenarios in the dimensions of challenge, trainability, safety margin, manipulation load, and program compliance, and form a comprehensive score that can be used for optimization and screening.

[0065] Four core evaluation indicators are set: flight quality deviation indicators, safety margin indicators, control requirements and load indicators, and training procedure compliance indicators.

[0066] Flight quality deviation indicators are used to measure whether an aircraft deviates from its target state, such as flight path, altitude, speed, and attitude. The indicator calculation does not consider single-point errors, but rather the integral of the error, cumulative time, peak value, and persistence, avoiding misjudgments based on instantaneous spikes.

[0067] Safety margin metrics are used to measure whether a system is approaching or exceeding safety boundaries, such as envelope margin, number of exceedances, critical thresholds, load limits, speed limits, and attitude limits. Strong penalties should be applied to exceeding the safety margin, while scenarios approaching the safety margin but not exceeding it should be given higher training value weights, encouraging the system to generate challenging scenarios that are close to the safety margin but manageable.

[0068] Maneuver requirements and load indicators are used to measure the intensity of maneuvering required to complete a task. They can be used to calculate the rate of change in maneuvering input, the number of directional reversals, the proportion of continuous high-intensity maneuvering, and the frequency of automatic / manual takeover. These indicators can distinguish between dangerous scenarios that do not require maneuvering and scenarios that require continuous and precise maneuvering.

[0069] Training procedure compliance metrics are used to measure whether a scenario triggers training objectives, such as the occurrence and duration of alarms, the number of mode switches, whether unstable approach criteria are met, and whether certain training process nodes are entered. This dimension allows the scenario to generate a more realistic flight training syllabus, rather than purely physical limits.

[0070] To maintain consistency and comparability across different device models, training phases, and initial conditions, the metric calculation follows three strategies: normalization, time window definition, and event analysis and statistics.

[0071] Normalization enables unified comparison and weighting of indicators with different dimensions and intervals. For example, normalization can be performed by target value or envelope margin.

[0072] The time window is defined by dividing the analysis window into different flight phases. During the approach phase, the focus is on stability and glide control; during the landing phase, the focus is on vertical load and attitude protection; and during the taxiing phase, the focus is on directional control and braking efficiency.

[0073] Event analysis and statistics mark key events in the logs, such as alarm trigger and clearing time, mode switching time, start / end of exceeding limits, and start / end of key control actions. The duration and recovery speed are calculated for each segment.

[0074] After extracting and normalizing various indicators, a reward function system needs to be constructed to comprehensively determine the value of a scene. This system is used not only to quantitatively rank candidate scenes but also to constrain and guide the automatic generation algorithm, making the generated results more consistent with the training objectives and safety requirements. The comprehensive reward function expression is as follows: In the formula: This represents the overall reward value for a single scenario, used to quantify the training value of special scenarios. A higher value indicates a higher training value for the scenario. , , , , These represent the weighting coefficients corresponding to each sub-indicator and the penalty item, respectively. Indicators representing deviations in flight quality; Indicates a safety margin indicator; This indicates the manipulation of demand and load indicators; Indicates the compliance index of the training procedure; This represents the out-of-bounds / untrainable penalty. It takes the value of 1 when a scenario involves a safe out-of-bounds or untrainable situation, and takes the value of 0 when there are no abnormal situations. This makes the system tend to generate scenarios with high training value and challenges but still trainable.

[0075] The reward function uses a single scene as the basic evaluation unit and usually performs a unified settlement after the scene ends. It can comprehensively consider multiple aspects such as the occurrence of anomalies, the handling process, and the final result, and avoid evaluation bias caused by local instantaneous performance.

[0076] The reward function is designed to maximize training value while considering safety boundaries. This means it encourages reasonably challenging scenarios while strictly suppressing extreme cases that are clearly untrainable or exceed the safety envelope. Specifically, factors such as flight quality deviations, control requirements, and program triggers, provided controllability is maintained, are typically reflected in the reward as positive contributions, reflecting their training value. Conversely, exceeding safety boundaries, severe structural or system overruns, and ultimately irreversible loss-of-control states are designed as strong penalties. Their occurrence significantly reduces the overall score, thus preventing the algorithm from generating high-risk but meaningless training scenarios. Ordinary scenarios with small deviations that do not trigger key training points receive a degree of negative correction to reflect their limited but still valuable training contribution.

[0077] Step 5: Construct an intelligent agent for generating special situation scenarios based on reinforcement learning. With the comprehensive reward function as the optimization objective, and combining the simulation environment, standardized description specifications for special situation scenarios, and sequence generation framework, construct a closed-loop automatic generation mechanism for special situation scenarios and output a batch of candidate special situation scenarios.

[0078] The system integrates scenario generation, simulation execution, result evaluation, and policy updates into a continuously operating closed-loop system. Simply having scenario generation capabilities is insufficient to generate training benefits. Only when simulation rounds can run automatically, data can be recorded in a standardized manner, evaluation metrics can be generated stably, and these metrics can be continuously fed back to the policy optimization module, can the system achieve continuous improvement and scalable generation capabilities.

[0079] A special scenario generation agent is constructed based on the Proximal Policy Optimization (PPO) algorithm. This algorithm is adaptable to a mixture of discrete and continuous action spaces, exhibits strong training stability and high sample utilization, and is suitable for the sequential decision-making scenarios in this scheme. The core design of the agent includes three parts: state space, action space, and network structure.

[0080] The agent's input is a concatenated one-dimensional state vector, composed of four types of sub-vectors: flight state sub-vector, control input and change sub-vector, training phase and time sub-vector, and situation history and difficulty estimation sub-vector. The flight state sub-vector includes altitude, airspeed, angle of attack, sideslip angle, and attitude angle. The control input and change sub-vector includes control inputs such as elevator, aileron, rudder, and thrust, as well as the rate of change of control. The training phase and time sub-vector includes flight phase markers and phase durations. The situation history and difficulty estimation sub-vector includes the number of injected situations, situation codes, and scenario difficulty level.

[0081] The agent's output is a hybrid discrete and continuous action vector, fully specifying the next situation element, corresponding one-to-one with the situation quadruple description specification. The discrete part includes object selection, situation type selection, injection method selection, and triggering method category selection. The continuous part includes intensity parameters and temporal parameters. Intensity parameters include bias amplitude, time constant, and drift rate, while temporal parameters include trigger time, trigger height, duration, and change slope. The agent adopts an actor-critic network structure. The actor network is responsible for outputting the action policy, and the critic network is responsible for evaluating the value of the current state. Both networks are constructed using three-layer fully connected neural networks, with the number of hidden layer neurons being 256, 128, and 64 respectively. The activation function is a linear rectified function, and the output layers correspond to the action probability distribution and state value estimation, respectively.

[0082] In closed-loop scenario generation, a control reference needs to be introduced during the simulation execution phase to complete the flight mission under given special conditions. Different control reference modes directly affect the scenario reproduction characteristics and evaluation results. Therefore, three switchable execution modes are set to adapt to different types of training and evaluation needs.

[0083] The first type is the automatic control logic mode. This mode executes flight missions based on pre-set automatic control methods. It features a high degree of standardization, strong controllability, and ease of batch operation, making it suitable for basic evaluation of large-scale automatically generated scenarios.

[0084] The second type is the historical control data playback mode. This mode plays back real flight control records and compares the impact of different special situation configurations on flight results under the same control input conditions, which can provide strong comparability for verification and source tracing analysis.

[0085] The third mode is a hybrid mode. In this mode, historical manipulation playback is the primary execution method. When the system detects a clear uncontrollable trend or an impending boundary violation, it automatically switches to safety protection logic or auxiliary control strategies to avoid meaningless crashes due to playback mismatch. This method can both preserve the authenticity of historical manipulation and provide safety protection when necessary, thereby effectively improving the utilization efficiency of simulation samples.

[0086] In closed-loop mode, the scene generation and training process in each round is executed in a fixed order.

[0087] First, the agent outputs a complete sequence of special situations based on the current policy, or generates individual special situation elements step by step.

[0088] The system then writes the generated objects, types, model methods, and timing parameters into the scenario injection plan file, and simultaneously generates random seeds to ensure reproducibility.

[0089] The simulation engine then executes the flight mission according to a unified control reference. The system injects special situations as planned at preset times or under preset conditions, and continues to evolve.

[0090] Throughout the round, the system continuously records flight status quantities, control inputs, system modes, alarm signals, and emergency injection events and timestamps.

[0091] After the simulation, the round data is evaluated and scored, and out-of-bounds states, untrainable states, or numerical anomalies are marked.

[0092] The current round's sample is then used as input for the policy update module, which can be implemented based on any reinforcement learning or rule optimization framework.

[0093] The above process continues to cycle, gradually forming a large-scale set of candidate scenarios and a strategy optimization trajectory.

[0094] Step 6: Conduct quantitative evaluation of the training value of candidate special scenarios, similarity analysis and redundancy elimination, and screen out special scenarios with high training value.

[0095] To transform the large number of automatically generated candidate scenes into structured and reusable training scenes, it is necessary to introduce mechanisms such as similarity analysis, redundancy removal, and hierarchical management to systematically screen and organize the scenes.

[0096] First, based on multi-dimensional comprehensive evaluation indicators and a comprehensive reward function, a unified quantitative evaluation is conducted on all candidate special situation scenarios. The simulation data for each group of candidate special situation scenarios is preprocessed, including data alignment, missing value completion, and outlier filtering. Analysis windows are divided according to flight phases. Combined with key event statistics, the comprehensive reward value for each group of scenarios is calculated, completing the quantitative ranking of training value. Low-value scenarios with comprehensive reward values ​​below a preset threshold are eliminated, while high-value candidate scenarios with reward values ​​above the threshold are retained. In this embodiment, the preset threshold is set to 0, eliminating low-value scenarios with comprehensive reward values ​​below 0.

[0097] To identify redundant and similar scenarios, a similarity evaluation mechanism is established from two levels: generation parameters and simulation results.

[0098] The first category is parameter similarity, which is based on the structural information of the incident sequence, including elements such as the anomalous object, anomaly type, trigger time window, intensity parameter, and evolution mode. By constructing a weighted distance metric for the above parameters, the similarity between two scenarios at the design level can be quickly determined. This method has low computational complexity and is suitable as a basis for preliminary screening and clustering. In the weighted distance of parameter similarity, the weights are: incident object 0.3, type 0.3, intensity parameter 0.2, and trigger time 0.2.

[0099] The second category is result similarity, which is based on simulation output data. It aligns and compares time-series curves such as track deviation, speed change, attitude response, key alarms, and mode switching sequences, and evaluates their similarity through statistical feature matching. This method directly reflects the flight experience and training effect, and is suitable for fine-tuning.

[0100] Scenes are clustered into coarse groups based on parameter similarity, and representative samples are determined within each group based on result similarity, thereby achieving a balance between computational cost and screening accuracy.

[0101] When two or more scenes are judged to be highly similar within a certain threshold range, no samples are directly deleted. Instead, a representative selection strategy is implemented, retaining representative samples with higher training value from the same type of scenes based on the principle of maximizing reward value. After the representative sample selection is completed, a unified identifier is assigned to each class of similar scenes, and this identifier is written into the scene record for subsequent tracking, expansion, and version management.

[0102] To ensure that the training process has a progressive nature from simple to complex, a difficulty grading mechanism based on computable rules is introduced. The scenarios are automatically divided into three progressive difficulty levels: basic, intermediate, and advanced, based on dimensions such as structural complexity, safety margin, and manipulation load.

[0103] The characteristics of basic-level scenarios include a small number of special situations, usually single special situations or weakly coupled combinations, a large safety margin, a clear distance from the boundary threshold, and a low to medium level of maneuvering load during task handling. They are mainly used for basic training and the initial familiarization phase.

[0104] The characteristics of the advanced scenario include the presence of two to three superimposed special situations, requiring obvious attitude corrections or procedural handling. The safety margin is reduced, but there is still a clear recovery space, which can be used to train the ability to identify and handle complex situations in a standardized manner.

[0105] Advanced scenarios are characterized by strong coupling issues that evolve over time, such as gradual deterioration or intermittent occurrence. The system state is close to the training threshold but still within manageable limits, placing high demands on energy management, pattern management, and fine manipulation. They are primarily used for high-level training and evaluation.

[0106] The aforementioned hierarchical strategy can be used not only for training course design but also as a constraint on the generation algorithm. For example, in the early stages of system training, it can be restricted to generating only basic-level scenes to avoid generating a large number of untrainable samples. As the strategy matures, advanced and high-level scenes can be gradually added, thus forming a stable and controllable scene evolution path.

[0107] In a preferred embodiment of the present invention, the method further includes step seven: storing high-training-value special situation scenarios into a special situation scenario library to realize the reuse, management and dynamic optimization of special situation scenarios.

[0108] The selected high-value scenarios are transformed into training assets that can be managed and reused long-term, and are stably integrated into the training process. Unlike simple file collections, the scenario library is built as a system resource with retrieval, version management, reproduction control, and continuous evaluation capabilities, thereby supporting flexible access and dynamic maintenance at different stages and in different training tasks.

[0109] To ensure the integrity and traceability of the scenarios entered into the database, each scenario record is described according to a unified data structure, including six core information categories: basic metadata, scope of application description, special situation description information, difficulty and training target labels, simulation example files, evaluation and result summary.

[0110] Basic metadata includes information such as scene identifier, version number, creation time, applicable machine type, configuration, and system software version, which is used to support asset management and consistency control.

[0111] The scope description is used to clarify the applicable flight phases for this scenario, such as takeoff, climb, cruise, approach, and go-around, as well as the corresponding training subjects, to ensure correct matching in the course arrangement and application process.

[0112] The incident description information records the complete incident sequence quadruple parameters, including the object, type, injection method, and timing information. It also saves the random seed used in this simulation and the mutual exclusion rule check results to ensure that the subsequent reproduction is consistent with the original environment.

[0113] Difficulty and training objective labels include scenario difficulty levels, basic / advanced / advanced, and corresponding training objectives, such as energy management, pattern management, fault handling, or stability control, to support search, filtering, and personalized recommendations.

[0114] The reproduction example files include scenario plan files, initial condition files, and necessary environment or system configuration snapshots to ensure consistency in reproduction across versions and platforms.

[0115] The assessment and results summary stores key assessment metrics, overall reward values, timestamps of important events, and optional summaries of typical response curves for easy analysis and archiving.

[0116] To adapt to changes brought about by system model or environment configuration updates, a version management mechanism is introduced into the scene library. The same scene identifier can generate a new version record as the model changes, while retaining historical versions, thereby ensuring training consistency and traceability.

[0117] The completed scenario library can be used in various training scenarios, including three core application modes: training course arrangement mode, assessment mode, and personalized recommendation mode.

[0118] The training course arrangement mode is based on difficulty levels. The system can organize scenarios into a progressive course structure. The basic level is used to establish standard operating procedures, the intermediate level is used to enhance comprehensive handling capabilities, and the advanced level is used for training close to the threshold. During arrangement, specific types of scenarios can be selected based on tags, such as high-load, strongly coupled, or frequently switching scenarios, for combination configuration.

[0119] For assessment-type tasks, the evaluation model can select a fixed set of scenarios as a unified evaluation sample, and ensure consistency of assessment conditions for each assessment through a scenario reproduction mechanism. The system can provide scoring references based on stored indicators, enabling horizontal comparability among different trainees.

[0120] The personalized recommendation system can analyze a trainee's weaknesses in areas such as deviation control, manipulation load, or response time based on their historical training results, and then match them with scenarios that have corresponding tags. Through targeted recommendations, a customized training path for individual skill improvement is formed.

[0121] To ensure the scenario library continuously aligns with actual training conditions, a data-closed-loop mechanism based on pilot performance and feedback is introduced to dynamically calibrate the scenario library. During training, the system automatically records trainee operational data and collects subjective evaluations regarding perceived difficulty, realism, and training effectiveness.

[0122] These data will be used to reverse-calibrate the evaluation system. When a certain type of scenario consistently achieves high scores but is generally considered to have limited training value, the system can appropriately adjust the weights of relevant metrics or label definitions. When a certain type of scenario is widely reported as too difficult or untrainable, its screening threshold can be increased or additional constraints can be imposed on the generation stage.

[0123] Meanwhile, as the system model is updated or the training objectives change, the scene library will periodically re-execute the screening and classification process, gradually forming a dynamic evolution mechanism of generation, screening, storage, training, feedback, and regeneration, so that the scene library can maintain its applicability and coverage over time.

[0124] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0125] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for automatically generating special scenarios during flight training, characterized in that, Includes the following steps: Step 1: Construct an integrated flight simulation operation environment for training and evaluation, and build a unified and standardized special situation injection interface to obtain the basic carrier for special situation scenario simulation operation and injection. Step 2: Establish a multi-dimensional parameterized description model and hierarchical parameter space for special situation scenarios. Represent a single special situation element in a structured way as a unified quadruple containing the target object, special situation type, injection method, and timing parameters, thereby obtaining a standardized description specification for special situation scenarios. Step 3: Model the special situation scenario generation process as a multi-step sequential decision-making process to obtain a standardized generation framework for the special situation scenario sequence; Step 4: Construct a multi-dimensional comprehensive evaluation index for training effectiveness, and based on the evaluation index, form a comprehensive reward function to quantify the training value of special scenarios, thereby obtaining the quantitative standard for the training value of special scenarios and the optimization goal of reinforcement learning. Step 5: Construct an intelligent agent for generating special situation scenarios based on reinforcement learning. With the comprehensive reward function as the optimization objective, and combining the simulation environment, the standardized description specification of special situation scenarios, and the sequence generation framework, construct a closed-loop automatic generation mechanism for special situation scenarios and output a batch of candidate special situation scenarios. Step 6: Conduct quantitative evaluation of the training value of candidate special scenarios, similarity analysis and redundancy elimination, and screen out special scenarios with high training value.

2. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The steps for constructing the basic carrier for the simulation and injection of the special situation scenario are as follows: The integrated flight simulation environment integrates a flight dynamics model adapted to all flight phases, an aircraft system model with an injectable interface, an environment model that supports time-series evolution, a multi-mode human-machine interaction interface, and a full-process data acquisition and recording module. The established special situation injection interface supports standardized structured input of special situations, injection of multiple types of special situation evolution modes, multi-condition combination triggering and recovery control, injection conflict and mutual exclusion rule verification, and also has the ability to trace the entire special situation injection process and reproduce scenarios.

3. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The steps for constructing the standardized description specification for the special situation scenario are as follows: A multidimensional parameterized description model is constructed using a hierarchical modeling approach, which sequentially sets up an object layer, a type layer, an injection method layer, and a timing layer. The object layer limits the scope of the special situation, the type layer defines the anomaly category of the special situation, the injection method layer clarifies the implementation model of the special situation in the simulation environment, and the timing layer defines the triggering, evolution, and recovery rules of the special situation. Finally, a single situation element is structured as a unified four-tuple structure containing the target, situation type, injection method, and timing parameters.

4. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The steps for constructing the hierarchical parameter space are as follows: Set up three types of parameter sets: system special situation parameter set, environmental special situation parameter set, and timing parameter set; The system incident parameter set is used to describe the abnormal behavior and performance degradation characteristics of various aircraft systems, the environmental incident parameter set is used to describe the abnormal and disturbance characteristics of external flight conditions, and the time series parameter set is used to describe the triggering, duration, evolution, and recovery rules of incidents in the time dimension.

5. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The execution steps of the standardized generation framework for the special situation scenario sequence are as follows: The complete sequence of special situation scenarios is represented as a sequence of multiple special situation elements in chronological order, and a single-step decision corresponds to the generation of a single special situation element; The single-step decision-making process sequentially executes four stages: state reading, candidate generation, constraint filtering, and selection and recording. State reading is used to obtain the current flight status, system mode, and incident history. Candidate generation is used to generate candidate incident elements from the hierarchical parameter space. Constraint filtering is used to eliminate candidates that do not meet physical rules and training requirements. Selection and recording is used to determine the final injected incident elements and complete the structured information retention.

6. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The quantitative standard for the training value of the special scenario and the steps for constructing the reinforcement learning optimization objective are as follows: Four core evaluation indicators are set: flight quality deviation indicators, safety margin indicators, control requirements and load indicators, and training procedure compliance indicators. Among them, the flight quality deviation index is used to measure the cumulative error and persistence of the aircraft's state deviating from the target state; the safety margin index is used to measure the degree to which the flight process approaches or exceeds the safety boundary; the control requirements and load index is used to measure the control intensity required to complete the flight mission; and the training procedure compliance index is used to measure the degree to which the scenario triggers the key points of the training outline. A comprehensive reward function with out-of-bounds penalty terms is constructed based on four core indicators, which serves as the optimization objective for reinforcement learning.

7. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The steps for constructing the intelligent agent for generating special scenarios are as follows: The input to the agent is a concatenated state vector, which contains four types of information: flight state, control input, training phase, special situation history, and difficulty estimation. The agent's output is a mixed discrete and continuous action vector, where the discrete part is used to determine the target, type, injection method, and triggering method of the special situation, and the continuous part is used to determine the intensity parameters and timing parameters of the special situation; The agent continuously optimizes the special situation generation strategy through reinforcement learning algorithms, guided by the comprehensive reward function.

8. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The single-round execution steps of the closed-loop automatic generation mechanism for special scenarios are as follows: The special situation scenario generation agent outputs a special situation scenario sequence according to the current strategy, writes it into a standardized scenario injection plan file, and generates random seeds to achieve scenario reproducibility. The simulation environment executes the flight mission according to the preset switchable control reference mode, and completes the special situation injection and full-process simulation evolution according to the injection plan; The entire simulation process includes collecting and recording all data, such as flight status, control inputs, system modes, alarm signals, and emergency injections. After the simulation, the evaluation index is calculated and the comprehensive reward value is settled. Invalid scenarios that are out of bounds, untrainable, or have abnormal values ​​are marked simultaneously. Based on the valid samples of this round, the agent's strategy is optimized, and batch special scenarios are automatically generated through repeated iterations.

9. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The quantitative evaluation steps for the training value of the candidate special situation scenarios are as follows: Based on multi-dimensional comprehensive evaluation indicators, the simulation operation data of each candidate special situation scenario is normalized, the analysis window is divided according to the flight stage, and the comprehensive reward value of a single scenario is calculated by combining the statistical results of key events. The training value is then quantitatively ranked, and low-value scenarios with comprehensive reward values ​​lower than the preset threshold are eliminated.

10. The method for automatically generating special flight training scenarios according to claim 1, characterized in that, The steps for similarity analysis, redundancy removal, and difficulty grading of the candidate special situation scenarios are as follows: Similarity analysis is conducted from two dimensions: generated parameters and simulation results. Parameter similarity is based on the structural information of the special situation sequence to construct a weighted distance metric to complete coarse clustering, while result similarity is based on the time series data of the simulation output to complete fine screening through statistical feature matching. Based on the similarity analysis results, redundant scenarios are eliminated and representative samples with the highest training value among similar scenarios are retained. Based on the structural complexity, safety margin, and operational load of the scenarios, the selected special situations are divided into three progressive difficulty levels: basic, intermediate, and advanced.

Images