Training priority determination method and system based on statistical analysis of aviation event data, electronic device and storage medium

Abstract

The application relates to the technical field of aviation safety and flight training data analysis, and provides a training priority determination method and system based on aviation event data statistical analysis, an electronic device and a storage medium, which solve the problems of poor precision and poor risk pertinence of flight training subject design. The method comprises the following steps: extracting aviation unsafe event data of a plurality of heterogeneous data sources, and extracting threat, error and non-expected aircraft state to determine the crew management state; adopting group decision-making combined with a distributed scoring node and a blockchain consensus mechanism to perform three-level scoring on various threats from the aspects of training effect, occurrence probability and consequence severity, so as to form a threat scoring matrix; statistically determining the crew management rate and actual occurrence frequency of each threat, calculating a comprehensive risk value, sorting all threats to determine the training priority, and embedding the threats with high priority and related errors into the scene design of the flight training subject. The application improves the precision of flight training subject design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aviation safety and flight training data analysis, and in particular to a training priority determination method and system based on aviation event data statistical analysis, an electronic device and a storage medium. BACKGROUND

[0002] In the field of aviation safety management, in order to improve the emergency disposal and risk response ability of the flight crew, a method is needed that can scientifically and automatically identify the risk types and disposal short boards that should be prioritized for strengthening in the current training system according to historical unsafe event data. This method aims to accurately invest limited training resources into the most safety-impacting and training-urgent aspects, thereby systematically improving overall safety.

[0003] There is currently a risk ranking scheme based on a two-dimensional matrix of event frequency and consequence severity. This scheme first standardizes the classification of various safety events, then counts the number of times each type of event occurs in historical data, and combines expert group qualitative assessment of the severity of the consequences that events may cause, and finally locates in the preset risk matrix to determine the risk level of different events and accordingly suggests training priorities.

[0004] This scheme has limitations in risk dimension coverage, mainly relying on the objective frequency of event occurrence and the severity of consequences, while ignoring the impact of crew performance differences in event disposal on risk assessment. At the same time, the expert evaluation process is easily disturbed by subjective cognitive differences, and the formation of consensus lacks transparent and traceable mechanism guarantee, which may cause deviation between risk assessment results and actual training needs. SUMMARY

[0005] The present application provides a training priority determination method and system based on aviation event data statistical analysis, an electronic device and a storage medium to solve the problem of poor precision and poor risk targeting of flight training subject design in the prior art.

[0006] To solve the above technical problems, in a first aspect, the present application provides a training priority determination method based on aviation event data statistical analysis, comprising:

[0007] extracting aviation unsafe event data from a plurality of heterogeneous data sources;

[0008] extracting the threat, error and undesired aircraft state corresponding to each aviation unsafe event from the aviation unsafe event data, and determining the management state of the crew to the threat, error and undesired aircraft state, the management state being a managed state or an unmanaged state;

[0009] The threat scoring matrix is generated by using a group decision algorithm, combining a distributed scoring node, and performing three-level scoring on each type of threat from three dimensions of training effect, occurrence possibility, and severity of resulting consequences under a blockchain consensus mechanism.

[0010] The statistical crew management rate and actual occurrence frequency of each type of threat are combined with the threat scoring matrix, the crew management rate, and the actual occurrence frequency of the threat to calculate a comprehensive risk value of each type of threat.

[0011] According to the comprehensive risk value, all threats are sorted, the training priority is determined based on the sorting result, and at least one threat with high training priority and the error associated with the threat are embedded into the scene design of the flight training subject.

[0012] Optionally, the extraction of the threat, error, and undesired aircraft state corresponding to each aviation safety event from the aviation safety event data and the determination of the management state of the crew to the threat, error, and undesired aircraft state include:

[0013] The aviation safety event data is input into a dynamic knowledge graph model constructed based on the threat and error management (TEM) classification method for entity recognition to extract the threat, error, and undesired aircraft state corresponding to each aviation safety event.

[0014] The management state of the crew to the threat, error, and undesired aircraft state is determined by a time series attention mechanism, and the management state is a managed state or an unmanaged state.

[0015] Optionally, the determination of the management state of the crew to the threat, error, and undesired aircraft state by the time series attention mechanism includes:

[0016] The threat, error, and undesired aircraft state corresponding to all aviation safety events are respectively taken as threat entities, error entities, and undesired aircraft state entities, and an event entity sequence is constructed in chronological order based on the entity identifiers corresponding to each entity.

[0017] The event entity sequence is input into a time series analysis model based on a pre-configured time series attention mechanism, and the correlation strength value between each threat entity in the event entity sequence and all crew operation entities appearing in a predetermined time period is calculated by a time series attention calculation module in the time series analysis model, wherein the crew operation entities include standard procedure response actions and pilot autonomous intervention actions.

[0018] The dynamic threshold screening module in the time series analysis model dynamically adjusts the screening threshold according to the type attribute of the threat entity, obtains an adjusted threshold, and screens candidate crew operating entities from all crew operating entities, wherein the candidate crew operating entities have an association strength value greater than the adjusted threshold with each threat entity;

[0019] The action effectiveness verification module in the time series analysis model performs compliance and utility evaluation on each candidate crew operating entity to determine whether the candidate crew operating entity is an effective management action for the corresponding threat, and obtains an effectiveness determination result;

[0020] The state labeling logic module in the time series analysis model labels the corresponding management state for each threat entity based on the effectiveness determination result, and derives and labels the same management state for the associated error entity and the non-desirable aircraft state entity based on the management state of the threat entity and the causal association relationship between the threat entity and the corresponding error entity and non-desirable aircraft state entity queried from the dynamic knowledge graph.

[0021] Optionally, the time series attention calculation module in the time series analysis model calculates the association strength value between each threat entity in the event entity sequence and all crew operating entities appearing within a preset time period, including:

[0022] Each entity identifier in the event entity sequence is converted into a numerical vector, wherein the vector corresponding to the threat entity is defined as a query vector, and the vector corresponding to the crew operating entity is defined as a key vector;

[0023] For each threat entity, all appearing crew operating entities within a fixed time window after the timestamp corresponding to the threat entity are retrieved;

[0024] The dot product between the query vector of the threat entity and the key vector of each candidate crew operating entity is calculated to obtain a similarity score;

[0025] The similarity score is input into the nonlinear transformation layer of the time series attention calculation module for processing to obtain a standardized association weight;

[0026] The standardized association weight corresponding to each candidate crew operating entity is weighted and summed with the value vector corresponding thereto, wherein the value vector is obtained by the linear transformation layer of the time series attention calculation module from the vector corresponding to the candidate crew operating entity;

[0027] The module length of the weighted sum result is taken as the association strength value between the threat entity and the corresponding crew operating entity.

[0028] Optionally, the utilization is based on a group decision algorithm, combined with a distributed scoring node, and under a blockchain consensus mechanism, to perform three-level scoring on each type of threat from three dimensions of training effect, occurrence probability, and severity of resulting consequences, to generate a threat scoring matrix, including:

[0029] According to the total number of threat types, a corresponding number of distributed scoring nodes are created, and each distributed scoring node is composed of a scoring group formed by flight operation experts, flight instructors, and risk analysts;

[0030] Each distributed scoring node scores the specified type of threat in the three scoring dimensions of training effect, occurrence probability, and severity of resulting consequences respectively through an improved Delphi method, to generate a node consensus scoring vector;

[0031] The node consensus scoring vectors generated by all distributed scoring nodes are submitted to a blockchain network, and a practical Byzantine fault tolerance consensus mechanism is used for multiple rounds of exchange and verification. When more than a preset proportion of nodes agree on the scoring of a specific dimension of a specific threat, the scoring of the specific dimension is recorded as a consensus score in the distributed ledger;

[0032] For each threat type, the consensus scores of the three dimensions of the threat type are extracted from the distributed ledger of the blockchain, and combined to form a scoring vector corresponding to the threat type;

[0033] All threat type scoring vectors are arranged in order of threat type identifiers to construct a threat scoring matrix.

[0034] Optionally, the scoring of the specified type of threat in the three dimensions of training effect, occurrence probability, and severity of resulting consequences respectively through the improved Delphi method to generate the node consensus scoring vector includes:

[0035] Obtain the scoring task initiated by the scoring group coordinator in the distributed scoring node for the specified type of threat, which contains three dimensions and multiple iteration rules of the improved Delphi method, and the multiple iteration rules contain an improved convergence condition;

[0036] During the execution of the scoring task, obtain the original anonymous scores and original anonymous score reasons for the three dimensions independently submitted by each role member of the scoring group based on the local knowledge base and historical evaluation data in the first round;

[0037] Based on the original anonymous scores and original anonymous score reasons of each dimension, calculate the statistical distribution value of each dimension score, including the interquartile range, and distribute the statistical distribution value and all role members' original anonymous score reasons as feedback information to all role members for reference in generating new anonymous scores and new score reasons for each dimension.

[0038] recomputing statistical distribution values of the dimensions scores based on the new scores and the new score reasons, the new scores being revised anonymous scores or original anonymous scores, the new score reasons being revised anonymous score reasons or original anonymous score reasons;

[0039] repeating the distributing operation, the obtaining operation and the recomputing operation until one of the following improved convergence conditions is met: quartile ranges of the scores of the dimensions are all less than or equal to a preset convergence threshold; a preset maximum iteration round is reached; a preset proportion of the role members indicates that the scores are no longer modified; after the improved convergence condition is met, taking an arithmetic mean value of the scores of the dimensions in the last round as a consensus score of the threat types by the corresponding distributed scoring nodes, and combining consensus scores of all the dimensions in a fixed order of the training effect, the occurrence probability and the severity to generate the node consensus score vector.

[0040] Optionally, the reference feedback information generates the new anonymous scores and the new score reasons of the dimensions, including:

[0041] comparing the anonymous scores of the three dimensions of the role members in the last round with the statistical distribution values in the last round in the feedback information to identify whether there is a target dimension based on a comparison result, the target dimension including: a dimension whose anonymous score is outside the consensus interval and a dimension whose quartile range is greater than a preset divergence threshold;

[0042] when the target dimension exists, performing score self-adaptive correction based on original anonymous score reasons of all the role members in the reference feedback information to obtain the new anonymous scores and the new score reasons of the dimensions.

[0043] In a second aspect, the present application provides a training priority determination system based on statistical analysis of aviation event data, comprising:

[0044] an extraction module configured to extract aviation unsafe event data from a plurality of heterogeneous data sources;

[0045] a determination module configured to extract threats, errors and undesired aircraft states corresponding to each aviation unsafe event from the aviation unsafe event data, and determine management states of the threats, the errors and the undesired aircraft states by the crew, the management states being managed states or unmanaged states;

[0046] a scoring module configured to perform three-level scoring on each type of threat from three dimensions of training effect, occurrence probability and severity of resulting consequences by using a group decision algorithm in combination with distributed scoring nodes under a blockchain consensus mechanism, and generate a threat score matrix.

[0047] The computing module is configured to count the crew management rate and the actual occurrence frequency of each type of threat, and calculate a comprehensive risk value of each type of threat in combination with the threat scoring matrix, the crew management rate and the actual occurrence frequency of the threat.

[0048] The determining module is configured to sort all the threats according to the comprehensive risk value, determine a training priority based on the sorting result, and embed at least one threat with a high training priority and an error associated with the threat into a scene design of a flight training subject.

[0049] In a third aspect, the present application provides an electronic device, comprising:

[0050] A memory configured to store a computer program;

[0051] A processor configured to execute the computer program to implement the steps of the training priority determination method based on statistical analysis of aviation event data according to the first aspect.

[0052] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executable by a processor to implement the steps of the training priority determination method based on statistical analysis of aviation event data according to the first aspect.

[0053] The technical scheme provided by the present application has the following beneficial effects:

[0054] The present application realizes the unified collection of security data dispersed in different systems and formats, lays a data foundation for comprehensive analysis, converts unstructured text event reports into structured and quantifiable security elements, accurately identifies the handling performance of the crew in the event chain, combines the objective occurrence of threats, the subjective management ability of the crew and expert scoring, forms more comprehensive and three-dimensional risk assessment indicators, and reduces subjective bias. The analysis results are directly converted into specific training improvement measures to ensure that training resources are concentrated on the most critical safety short board, and the training is more targeted and effective.

[0055] Further, the present application first inputs aviation unsafe event data into a dynamic knowledge graph model based on the threat and error management classification method to identify and extract threats, errors and unexpected aircraft state entities in the event; then, the time sequence attention mechanism is used to analyze the correlation between crew operation and threats in the event sequence, so as to determine the management state of the crew to these security elements.

[0056] And, the technical solution realizes automatic and structured extraction of safety elements in event reports, and can accurately evaluate the management performance of the crew in a dynamic event sequence, thereby providing a reliable and refined analysis basis for subsequent risk quantification and training priority determination. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0058] Figure 1 A flowchart of a training priority determination method based on aviation event data statistical analysis provided by the embodiments of the present application;

[0059] Figure 2 A specific implementation schematic diagram of a training priority determination method based on aviation event data statistical analysis provided by the embodiments of the present application;

[0060] Figure 3 A structural schematic diagram of a training priority determination system based on aviation event data statistical analysis provided by the embodiments of the present application. DETAILED DESCRIPTION

[0061] The existing risk ranking scheme based on event frequency and consequence severity has limitations in guiding the setting of flight training priority. This scheme mainly relies on the objective statistics of historical events and the qualitative evaluation of experts on consequences, and the risk evaluation dimension does not include the actual management performance of the crew in the event handling process. At the same time, the formation process of expert consensus relies on traditional discussion, and its subjectivity and transparency still have room for improvement, which may cause deviation between the final risk ranking result and the real training demand.

[0062] To solve the above problems, the present application provides a training priority determination method based on aviation event data statistical analysis, introduces the key dimension of crew management performance, and uses a transparent and tamper-proof consensus mechanism for multi-dimensional evaluation, thereby constructing a more three-dimensional and objective risk evaluation model, so as to ensure that the ranking result of training priority can truly reflect the safety short board in the scene operation and the training demand, and effectively overcome the shortcomings of the existing scheme in evaluation dimension and process reliability.

[0063] For the person skilled in the art to better understand the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0064] The core of the present application is to provide a training priority determination method based on aviation event data statistical analysis, and a specific embodiment of the method is shown in the flowchart as Figure 1 The method comprises the following steps:

[0065] Step 101: Extracting aviation unsafe event data from multiple heterogeneous data sources.

[0066] In step 101, the multiple heterogeneous data sources refer to aviation safety data sources with different formats and structures, such as text format event reports, database format flight data records, and table format maintenance records, etc. Aviation unsafe event data refers to all relevant event record information that may affect or has affected flight safety. For any event, the present application embodiments can provide event number, event occurrence date, aircraft model, airport, flight phase, event label, event characterization, etc.

[0067] In the present application embodiments, first, data files from different sources and with different formats are accessed and read, then the original data is converted in format and aligned in field through data parsing technology, and is processed uniformly into a standardized intermediate data format, next, the converted data is cleaned to remove errors or duplicate entries, and finally, all the cleaned valid data is merged into a complete and unified aviation unsafe event data set, ready for the subsequent analysis steps.

[0068] For example, the system automatically captures all flight operation related event description information in the past year from three heterogeneous data sources, including A airline's text type event report library, B regulatory agency's database type safety report, and C airport's tabular ground support record. The system first converts all these different format data into a unified JSON intermediate format, then automatically filters out invalid entries with incomplete or obviously contradictory records, such as a report lacking specific timestamp, and finally integrates a standard and complete aviation unsafe event data set for the next step of analysis.

[0069] Step 102: extracting the threat, error and undesired aircraft state corresponding to each aviation unsafe event from the aviation unsafe event data, and determining the management state of the crew to the threat, error and undesired aircraft state, the management state being a managed state or an unmanaged state.

[0070] In step 102, the threat refers to an external condition or event that exists before the operation of the pilot in the aviation operation environment, which may pose a potential risk to flight safety. The error refers to the operation or judgment error that may cause the system state to deviate from the expected state during the execution of the task by the pilot. The undesired aircraft state refers to a non-planned and non-standard flight or ground state of the aircraft caused by the influence of the threat or error. The management state is used to identify whether the crew has successfully prevented the threat or error from developing into a more serious unsafe consequence through effective operation intervention, and the state is divided into managed and unmanaged.

[0071] In the embodiments of the present application, a unified event data set is first input into a dynamic knowledge graph model based on the threat and error management classification method, which can automatically identify the threat entity, error entity and undesired aircraft state entity described in the text. Subsequently, the system uses a time sequence attention mechanism to analyze a series of operation actions performed by the crew within a specific time window after the threat occurs, and checks the compliance and effectiveness by calculating the correlation strength between the operations and the threat, and combining the pre-set standard operation procedure library, so as to intelligently judge whether the crew has effectively managed the current threat. Finally, the system automatically infers and labels the management state of the associated error and undesired aircraft state according to the management state determination result of the threat and the causal relationship between the entities in the knowledge graph.

[0072] For example, based on the data set obtained in the previous step, the system uses the dynamic knowledge graph model to accurately identify the threat entity "low-level wind shear", the error entity "initial energy management is slightly conservative", and the undesired aircraft state entity "aircraft below glide path" from an event report describing "flight encounters low-level wind shear during approach phase". Then, the time sequence analysis model finds that the crew performed the standard operation "immediately execute the missed approach procedure" after the "low-level wind shear" occurred, and the operation is completely compliant and can effectively respond to the threat, so it is determined that the management state of the threat is "managed". Based on the causal chain in the knowledge graph that "low-level wind shear leads to conservative initial energy management, which in turn causes the aircraft to be below the glide path", the system automatically labels the management state of the associated error and undesired aircraft state as "managed".

[0073] Step 103: Using a group decision algorithm, combined with distributed scoring nodes, under a blockchain consensus mechanism, a threat scoring matrix is generated by scoring each type of threat from three dimensions: training effectiveness, likelihood of occurrence, and severity of resulting consequences.

[0074] In step 103, the group decision algorithm is a decision-making method that aggregates the opinions of multiple experts to reach a collective consensus. The distributed scoring nodes refer to units composed of different expert groups that independently perform scoring work in the network. The blockchain consensus mechanism is a technical protocol that ensures all participants in a distributed network can agree on a certain data value and prevents data from being tampered with. The training effectiveness dimension is used to assess the potential value of improving the ability of unit management to manage this threat through targeted training. The likelihood of occurrence dimension is used to assess the probability of this type of threat occurring in actual operation. The severity of resulting consequences dimension is used to assess the severity of the safety consequences that may be triggered if the threat is not effectively managed. The three-level scoring refers to setting multiple levels (e.g. high, medium, low) for each dimension to quantitatively evaluate. The threat scoring matrix is a mathematical table, with rows representing different types of threats and columns representing the three scoring dimensions of training effectiveness, likelihood of occurrence, and severity of resulting consequences. Each element in the matrix is the consensus score value of the corresponding threat in the corresponding dimension.

[0075] In the embodiments of the present application, first, according to the total number of identified threat types, a corresponding number of distributed scoring nodes are formed, each node being composed of a scoring group containing a flight operations expert, a flight instructor, and a risk analyst. Each node uses the improved Delphi method to independently submit scores and reasons from the three dimensions of training effectiveness, likelihood of occurrence, and severity in multiple rounds, with the group members being anonymous. The opinions are continuously revised by referring to the statistical feedback of the previous round until a group consensus is reached, forming the consensus scoring vector of the node. Subsequently, the consensus scoring vectors of all nodes are submitted to a blockchain network, and through the built-in practical Byzantine fault tolerance consensus mechanism, multiple rounds of exchange and verification are performed. When a sufficient proportion of nodes agree on the score of a specific dimension of a specific threat, the consensus score is permanently recorded on the distributed ledger. Finally, the system extracts the final consensus scores of all threat types in the three dimensions from the blockchain ledger and arranges them in order of threat type identifier to construct and generate the threat scoring matrix.

[0076] For example, assume that the system has identified multiple threats, such as "low-level wind shear", "communication interruption", etc. For this purpose, a plurality of distributed scoring nodes corresponding to the threats are formed. Among them, the scoring node for the "low-level wind shear" threat, after a plurality of rounds of anonymous scoring and reason elaboration by the experts in the team, finally reaches an internal consensus of the team on the threat in the "training effect" dimension (high value), "occurrence possibility" dimension (medium), and "severity" dimension (high), forming a vector. The group consensus vectors of all threat types are uploaded to the blockchain network, and after a plurality of rounds of verification among nodes, a global consensus is finally reached on the scoring of each dimension of all threats. The system then extracts these consensus scores from the blockchain to form a complete threat scoring matrix, in which the three dimensions of the row where "low-level wind shear" is located record the consensus results.

[0077] Step 104: Calculate the comprehensive risk value of each type of threat by combining the threat scoring matrix, the management rate of the crew, and the actual occurrence frequency of the threat.

[0078] In step 104, the management rate of the crew is the proportion of the number of events in which the crew has managed a certain type of threat to the total number of events in all related unsafe events. The actual occurrence frequency refers to the number of times a certain type of threat appears in all unsafe events within a certain period of time. The comprehensive risk value is a value obtained through a specific calculation formula, which comprehensively reflects the expert consensus score of the threat, the difficulty of its management (indirectly reflected by the management rate), and its frequency of occurrence, and is used for final risk ranking.

[0079] In the embodiments of the present application, first, based on the event data with labeled management status in step 102, the management rate of each type of threat is calculated and counted. At the same time, the actual occurrence frequency of each type of threat in the historical data is counted. Then, from the threat scoring matrix generated in step 103, the scores of each type of threat in the three dimensions of training effect, occurrence possibility, and severity of resulting consequences are read. Finally, these data are substituted into a pre-set comprehensive risk value calculation formula, which usually performs weighted or multiplied combination operations on the management rate (reflecting the effectiveness of management), the occurrence frequency (reflecting the exposure rate), and the scores of the three dimensions (reflecting the inherent properties of the threat and the training value), so as to calculate a comprehensive risk value representing the overall risk level of each threat.

[0080] For example, following the preceding steps, the system statistics find that, among all the events recorded as “low-level wind shear” threats, the proportion of successful management by the crew (i.e., the crew management rate) is a certain value, and the actual occurrence frequency of the threat in the statistical period is a certain value. At the same time, the consensus score of “low-level wind shear” in the three dimensions of training effect, occurrence probability, and severity is obtained from the threat scoring matrix. Assuming that the formula for calculating the comprehensive risk value is wherein represents the comprehensive risk value, represents the severity score, represents the occurrence probability score, represents the crew management rate, represents the actual occurrence frequency, represents the training effect score. Substituting the above-mentioned values into this formula, the comprehensive risk value R of the “low-level wind shear” threat can be calculated.

[0081] Step 105: According to the comprehensive risk value, all threats are sorted, the training priority is determined based on the sorting result, and at least one threat with high training priority and the error associated with the threat is embedded into the scene design of the flight training subject.

[0082] In step 105, sorting refers to arranging all threat types in order from large to small (or from small to large) according to the comprehensive risk value. The training priority is determined according to the sorting result, and the threat sequence that needs to be paid attention to and strengthened in subsequent flight training. The error associated with the threat refers to the pilot error identified by the knowledge graph model in step 102, which has a causal relationship with a specific threat. The scene design of the flight training subject refers to the construction of a training script, scenario, or simulation machine task used to simulate a real flight environment.

[0083] In the embodiments of the present application, first, the comprehensive risk values of all threat types calculated in step 104 are sorted from high to low. Then, according to this sorting list, the order of investment of training resources is determined, and generally the higher the comprehensive risk value of a threat, the higher its training priority. Next, the system will select the top several threats in the sorting, and according to the dynamic knowledge graph constructed in step 102, automatically find the pilot errors most commonly associated with these high-priority threats. Finally, these high-priority threats and their typical associated errors are designed as core elements into the specific subject scene of the flight training simulator, for example, intentionally setting the threat condition in the simulation flight task, and examining whether the student can avoid or correctly handle the associated error.

[0084] For example, after the system calculates the comprehensive risk values of all threats (including "low-level wind shear", "communication interruption", "engine failure", etc.), the system ranks them in descending order. Assuming that the comprehensive risk value of "low-level wind shear" is the highest, ranking first, it is determined as the highest training priority. Then, the system queries the knowledge graph and finds that the most common error associated with "low-level wind shear" is "initial energy management is slightly conservative". Therefore, in the next cycle of pilot retraining simulator subjects, the training management department specially designs a training scene of "encountering strong wind shear in the approach phase", and the core assessment point of this scene is whether the pilot can effectively avoid or timely correct the error of "initial energy management is slightly conservative" when facing the threat of "low-level wind shear", so as to safely complete the reflight or approach. When embedding high-priority threats and associated errors into scene design, these key competency elements need to be considered simultaneously, so as to ensure that the designed training scene not only reproduces specific threats and errors, but also exercises and evaluates the pilot's performance in these core competencies. Key competencies include knowledge application, procedure application, automated flight path management, manual flight path management, communication, leadership and team collaboration, decision-making, situational awareness, and workload management. Simulator training element design includes threat-related design, error-related design, and non-desired aircraft state-related design. When embedding at least one threat with high training priority and the error associated with the threat into the scene design of the flight training subject, it is further clarified that the "scene design" specifically includes three core elements: threat-related design, error-related design, and non-desired aircraft state-related design, so that the construction of the training scene is more structured and targeted.

[0085] The method can automatically identify key safety threats and crew handling shortfalls in operation by systematically collecting and analyzing multi-source aviation safety data, and can perform multi-dimensional and reliable risk assessment on threats by integrating expert group wisdom and objective operation data. Ultimately, the method can accurately locate training needs and directly convert the analysis results into highly targeted training content, thereby improving the efficiency of investment in flight training resources and the risk management ability of pilots, and effectively supporting the continuous improvement of aviation safety level.

[0086] To solve the problem of insufficient precision in extracting safety elements in aviation unsafe events and determining the management state of the crew, in some embodiments, step 102: extracting the threats, errors, and non-desired aircraft states corresponding to each aviation unsafe event from the aviation unsafe event data, and determining the management state of the crew to the threats, errors, and non-desired aircraft states, as shown in Figure 2 , includes:

[0087] Step 201: input the aviation unsafe event data into a dynamic knowledge graph model based on the TEM taxonomy to perform entity recognition to extract threats, errors and undesired aircraft states corresponding to each aviation unsafe event.

[0088] In step 201, the TEM taxonomy is a framework theory for systematizing the classification of various risks in aviation operations. The dynamic knowledge graph model is a semantic network that can represent entities and their relationships. In this application, the nodes represent specific entities identified from the event data, such as threat entities, error entities, and undesired aircraft state entities, and the edges represent the causal or temporal relationships between these entities. The model can dynamically update its node and edge structures based on new input data. Entity recognition refers to the process of automatically identifying and extracting named entities of predefined types from unstructured text data.

[0089] In the embodiments of the present application, a knowledge framework is first constructed based on the TEM theory, which defines entity types such as threats, errors and undesired aircraft states and their mutual relationships. Then, the preprocessed aviation unsafe event text data is input into the model, which uses its internal entity recognition component to automatically scan the text and identify and extract threat entities, error entities and undesired aircraft state entities of predefined types. The model dynamically learns new associations between entities during the recognition process and updates the graph structure. The final output is a structured entity set corresponding to each unsafe event, which clearly identifies the threats present in the event, the errors caused and the undesired states caused.

[0090] Step 202: determine the management state of the crew for the threats, errors and undesired aircraft states through a temporal attention mechanism, the management state being a managed state or an unmanaged state.

[0091] In step 202, the time sequence attention mechanism is a calculation model that can pay attention to the importance of information at different positions in the time sequence. In this application, it is used to analyze which of the subsequent series of crew operation entities are truly relevant to the management of the threat entity after its appearance. There is a close cooperative relationship between the time sequence attention mechanism and the "dynamic knowledge graph model": the dynamic knowledge graph model is responsible for structurally extracting threat, error, and non-expected aircraft state entities and their causal relationships from aviation safety event data, and constructing a knowledge representation of the event; and the time sequence attention mechanism is based on the time sequence of the appearance of the entities, analyzes the time sequence correlation strength between the threat and the subsequent crew operation, and thus judges whether the crew has effectively managed the threat. The combination of the two realizes a complete analysis chain from static event structure to dynamic behavior management. The managed state indicates that the crew successfully mitigated or eliminated the risk through effective operation, and the unmanaged state indicates that the crew failed to effectively intervene.

[0092] In the embodiment of the present application, first, the entities extracted in step 201 are arranged in chronological order of the occurrence of the event to form an event entity sequence. The core of the time sequence attention mechanism is to analyze the correlation degree between all crew operation entities and the threat entity within a specific time window after the appearance of the threat entity. The mechanism calculates the correlation strength score of each threat entity with each crew operation entity in the window. Then, the system sets a dynamically changing filtering threshold according to the type of the threat, and filters out candidate crew operations whose correlation strength exceeds the threshold. Next, the candidate operations are checked in two aspects, one is to check whether they meet the prescribed operation procedures, and the other is to evaluate whether they are effective in mitigating or eliminating the current threat. Only the operations that are both compliant and effective will be determined as effective management actions for the threat. If an effective management action is found for a threat, its management state is marked as managed, otherwise as unmanaged. Finally, the system automatically derives the same management state for the errors and non-expected aircraft states associated with the determined threat according to the causal relationships between the threat, error, and non-expected aircraft state in the knowledge graph.

[0093] The following is a specific example:

[0094] After the system ranks the "low-level wind shear" as the highest training priority based on the comprehensive risk value, the training management department designs the scene of the flight training subject according to the analysis results output by the system, aiming at the threat and the associated error "initial energy management slightly conservative". In the next cycle of pilot retraining of A airline, a high-fidelity simulation scene named "encountering strong wind shear in the approach phase" is added to the simulation training subject. The core design elements of this scene are completely derived from the analysis conclusions of the method of the application, that is, the simulator dynamically simulates the threat entity of low-level wind shear during the approach of the aircraft, and focuses on monitoring the energy management strategy of the pilot in the early stage of the threat, to investigate whether he can effectively avoid or timely correct the specific error entity of initial energy management slightly conservative. The training evaluation standard is directly related to the judgment logic of the management state. If the pilot can timely, in compliance and effectively execute the standard operation such as the missed approach procedure, so as to successfully manage the threat and avoid the aircraft entering the non-desired aircraft state, it is determined that he passes the scene examination. This design ensures that the training content is highly consistent with the real safety risk and crew capability short board obtained through data analysis, and improves the pertinence and effectiveness of the training.

[0095] In the embodiments of the application, by combining the dynamic knowledge graph and the time sequence attention mechanism, the precise and automatic extraction of key safety elements in aviation unsafe events is realized, and an objective judgment of the management efficiency of the crew is made in accordance with the time sequence logic of the event, providing a high-quality and structured analysis basis for subsequent precise risk assessment and training priority ranking.

[0096] In order to further improve the automation and accuracy of judging the management state of the crew, in some embodiments, step 202: determining the management state of the crew to the threat, error and non-desired aircraft state by the time sequence attention mechanism, comprises:

[0097] Step 301: all aviation unsafe event corresponding threats, errors and non-desired aircraft states are respectively taken as threat entities, error entities and non-desired aircraft state entities, and based on the entity identifiers corresponding to each entity, an event entity sequence is constructed in chronological order.

[0098] In step 301, the entity identifier is an identity code uniquely representing an entity object. The time stamp is the time information recording the time when a certain entity occurs in the event. The event entity sequence refers to the sequence formed by arranging all the entities identified in an unsafe event in chronological order.

[0099] In the embodiments of the present application, the system first acquires all the entities that have been identified from the dynamic knowledge graph model, including threat entities, error entities and undesirable aircraft state entities. Then, the system reads the timestamp information attached to each entity. Next, the system arranges all the entities in the same unsafe event in chronological order from early to late according to the timestamp information, and finally forms a well-organized event entity sequence.

[0100] Step 302: input the event entity sequence into a time series analysis model pre-configured based on a time sequence attention mechanism, and calculate the correlation strength value between each threat entity in the event entity sequence and all the crew operation entities appearing within a preset time period through a time sequence attention calculation module in the time series analysis model, wherein the crew operation entities include standard procedure response actions and pilot autonomous intervention actions.

[0101] In step 302, the time sequence attention calculation module is a component specially used to calculate the strength of the correlation between different elements in a time sequence. The preset time period refers to a time range set for finding crew operations that may be related to a certain threat entity, usually starting from the time point when the threat entity appears. The correlation strength value is a numerical value used to quantify the closeness of the correlation between a threat entity and a crew operation entity in a time sequence.

[0102] In the embodiments of the present application, the system sends the event entity sequence generated in step 301 into the time series analysis model. The time sequence attention calculation module in the model starts to work, and it finds all the crew operation entities appearing within a fixed time length after each threat entity in the sequence. The module calculates the correlation between the threat entity and each crew operation entity found by a specific mathematical method, and calculates a numerical value representing the closeness of the correlation for each pair of relationships, i.e. the correlation strength value.

[0103] Step 303: through a dynamic threshold screening module in the time series analysis model, dynamically adjust the screening threshold according to the type attribute of the threat entity, obtain the adjusted threshold, and screen out candidate crew operation entities whose correlation strength values with each threat entity are all greater than the adjusted threshold from all the crew operation entities.

[0104] In step 303, the dynamic threshold screening module is a component that can automatically adjust the screening criteria according to different conditions. The screening threshold is a critical value used to determine whether the correlation strength value is large enough. Only the unit operation associated with the threat entity whose correlation strength value exceeds this critical value will be further considered. The type attribute of the threat entity refers to the specific category to which the threat belongs, such as a meteorological threat or a mechanical threat. The candidate unit operation entity refers to those unit operations whose correlation strength with the threat entity is large enough to be preliminarily screened out for subsequent detailed examination.

[0105] In the embodiments of the present application, the dynamic threshold screening module intelligently selects a suitable screening threshold according to the type of the threat entity being processed. This threshold may be different for different types of threats. Then, the module compares all the correlation strength values calculated in step 302 with the dynamically determined threshold, and only retains those unit operation entities whose correlation strength values are greater than the threshold. These retained operation entities are marked as candidate unit operation entities.

[0106] Step 304: Through the action effectiveness verification module in the time series analysis model, the compliance and effectiveness of each candidate unit operation entity are evaluated to determine whether the candidate unit operation entity is an effective management action for the corresponding threat, and an effectiveness determination result is obtained.

[0107] In step 304, the action effectiveness verification module is a component responsible for evaluating whether a unit operation is truly effective. Compliance evaluation is to check whether the operation conforms to the established and written operation procedures and standards. Effectiveness evaluation is to judge whether the operation can effectively cope with or resolve the threat associated with it in actual situations. The effectiveness determination result is a conclusion indicating whether the candidate unit operation entity is finally identified as an effective management action for the corresponding threat.

[0108] In the embodiments of the present application, the action effectiveness verification module performs two-step verification on each candidate machine group operation entity filtered out in step 303. The first step is to verify whether the operation is within the scope of the standard procedure, that is, whether it conforms to the regulations. The second step is to judge whether the execution of the operation can indeed play a role in managing the threat and reducing the risk in combination with the specific context of the threat. Only the operation that passes both steps of verification will be determined as an effective management action. The specific implementation process is as follows: first, based on the pre-set standard operation procedure library and expert rule library, the compliance of the candidate machine group operation entity is checked and matched to determine whether it conforms to the standard procedure; second, in combination with the threat type and the context, the effectiveness evaluation model (such as a statistical model based on historical data or a lightweight simulation deduction) is used to judge whether the operation can effectively alleviate or eliminate the threat; if the operation is both compliant and effective, it is determined as an effective management action. Example: when the “low-altitude wind shear” threat occurs, the candidate operation is “immediately execute the go-around procedure”, which is verified to conform to the standard procedure, and historical data shows that the operation can reduce the risk of loss of control caused by wind shear, so it is determined as an effective management action.

[0109] Step 305: Based on the effectiveness determination result, the state labeling logic module in the time series analysis model labels the corresponding management state for each threat entity, and based on the management state of the threat entity and the causal association relationship between the threat entity and the corresponding error entity and the non-desirable aircraft state entity queried from the dynamic knowledge graph, the management state of the associated error entity and non-desirable aircraft state entity is derived and labeled.

[0110] In step 305, the state labeling logic module is a component responsible for finally labeling the entity with a management state. The causal association relationship refers to the front and rear causal connection recorded in the dynamic knowledge graph, that is, the threat entity as the cause leads to the occurrence of the error entity, and the error entity as the cause leads to the occurrence of the non-desirable aircraft state entity.

[0111] In the embodiments of the present application, the state labeling logic module works according to the effectiveness determination result obtained in step 304. If a threat entity has at least one management action determined as effective, the management state of the threat entity is labeled as managed state, otherwise it is labeled as unmanaged state. Subsequently, the module queries the dynamic knowledge graph to find all error entities and non-desirable aircraft state entities directly or indirectly caused by the threat entity. Finally, the system sets the management state of these associated error entities and non-desirable aircraft state entities to the same management state as the threat entity at the root. After labeling the management state, the method further comprises: through a knowledge graph update interface, writing the labeled management state information back to the dynamic knowledge graph model as an entity attribute and storing it in association with the corresponding threat entity.

[0112] The following is a specific example:

[0113] The high-priority threat low-level wind shear and its associated error initial energy management slightly conservative identification is just due to the fine analysis process of a specific unsafe event in this step. In this event, the system first constructs the event entity sequence according to the entity identifier and the time stamp, and the order is the threat entity low-level wind shear, the error entity initial energy management slightly conservative, the undesired aircraft state entity aircraft below glide path and the crew operation entity executing the missed approach procedure. The time sequence attention calculation module in the time sequence analysis model then processes the sequence, which retrieves the crew operation entity executing the missed approach procedure within a preset time period after the threat entity low-level wind shear appears, and calculates a higher correlation strength value to quantify the close degree of the two in time sequence. Then, the dynamic threshold screening module dynamically selects a higher screening threshold according to the type attribute that the low-level wind shear belongs to the meteorological threat, and since the calculated correlation strength value exceeds this threshold, the crew operation entity executing the missed approach procedure is successfully screened as the candidate crew operation entity. Subsequently, the action effectiveness verification module checks the candidate crew operation entity, confirms that it fully complies with the provisions of the standard operating procedure and is highly effective in dealing with such wind shear threats as proven by historical data and expert knowledge, and thus determines it as an effective management action for the low-level wind shear threat. Finally, the state labeling logic module labels the management state of the threat entity low-level wind shear as the managed state based on this effectiveness determination result, and automatically derives and labels the same managed state for the associated error entity initial energy management slightly conservative and the undesired aircraft state entity aircraft below glide path based on the explicit causal relationship that the low-level wind shear leads to the initial energy management slightly conservative and in turn causes the aircraft below glide path, which is queried from the dynamic knowledge graph. This complete and automated determination chain provides key data support for subsequent accurate calculation of the crew management rate of this threat and finally locates it as the highest training priority.

[0114] In the embodiments of the present application, through the above-mentioned fine steps, the automated and accurate determination of the crew management state is realized, the determination process fully considers the time sequence of event occurrence, the logical association of operation and threat and the actual effectiveness of action, and the determination result is automatically transmitted to the related entities based on the causal relationship, thereby providing a highly reliable management state data foundation for subsequent risk assessment.

[0115] To further improve the accuracy of calculating the correlation strength between threat entities and crew operation entities, in some embodiments, step 302: the correlation strength value between each threat entity in the event entity sequence and all crew operation entities appearing within a preset time period is calculated by a time series attention calculation module in the time series analysis model, including:

[0116] Step 401: convert each entity identifier in the event entity sequence into a numerical vector, wherein the vector corresponding to the threat entity is defined as a query vector, and the vector corresponding to the crew operation entity is defined as a key vector.

[0117] In step 401, the entity identifier is a code that uniquely identifies an entity. The numerical vector is an ordered list of numbers that is used to mathematically represent an entity. The query vector is a numerical vector that specifically represents a threat entity and is used to actively initiate matching in calculations. The key vector is a numerical vector that specifically represents a crew operation entity and is used to be matched by the query vector.

[0118] In the embodiments of the present application, the time series attention calculation module first converts the identifier of each entity in the event entity sequence into a numerical vector of fixed length through an embedding layer. In this process, the numerical vector converted from all threat entities is specially defined as a query vector. And the numerical vector converted from all crew operation entities is specially defined as a key vector. At the same time, the vector corresponding to each crew operation entity will also generate another numerical vector called value vector through an independent linear transformation layer for subsequent steps.

[0119] Step 402: for each threat entity, retrieve all appearing crew operation entities within a fixed time window after the timestamp corresponding to the threat entity.

[0120] In step 402, the fixed time window refers to a pre-set duration of time calculated from the time of appearance of the threat entity. Retrieval refers to finding and identifying all eligible crew operation entities within this time range.

[0121] In the embodiments of the present application, the system iterates through each threat entity. For the threat entity currently being processed, the system defines a fixed length time range backward according to its recorded timestamp. Then, the system scans the event entity sequence within this time range to find all crew operation entities appearing within this time period. These found crew operation entities are called candidate crew operation entities.

[0122] Step 403: calculate the dot product between the query vector of the threat entity and the key vector of each candidate crew operation entity to obtain a similarity score.

[0123] In step 403, the dot product is a mathematical operation that calculates the degree of alignment of two vectors in a direction, and the result is a numerical value. The similarity score is the result of this dot product operation, and its numerical size represents the degree of similarity between the query vector of the threat entity and the key vector of the crew operation entity in the semantic space.

[0124] In the embodiment of the present application, for each candidate crew operation entity found in step 402, the time sequence attention calculation module takes out its key vector. At the same time, the query vector of the current threat entity is taken out. Then, the module calculates the dot product between the query vector and the key vector of the candidate crew operation entity. The dot product value obtained by this calculation is the initial similarity score between the current threat entity and the candidate crew operation entity.

[0125] Step 404: input the similarity score into the nonlinear transformation layer of the time sequence attention calculation module for processing to obtain the standardized association weight.

[0126] In step 404, the nonlinear transformation layer is a component that performs nonlinear mathematical conversion on input data. Standardization is a process of scaling a set of numerical values so that their sum is one. The standardized association weight is the similarity score after standardization, which represents the importance proportion of a certain specific operation relative to the current threat among all candidate operations.

[0127] In the embodiment of the present application, the time sequence attention calculation module collects the similarity scores of all candidate crew operation entities and the current threat entity obtained in step 403, inputs them into a nonlinear function for transformation, with the purpose of introducing complexity and stabilizing the numerical value. Then, the transformed scores are standardized to ensure that the sum of all scores is one. After processing, each candidate crew operation entity obtains an association weight between zero and one.

[0128] Step 405: weight and sum each candidate crew operation entity corresponding to the standardized association weight and the corresponding value vector, wherein the value vector is obtained by linear transformation of the vector corresponding to the candidate crew operation entity through the time sequence attention calculation module.

[0129] In step 405, the weighted sum is a mathematical operation that multiplies each value by a weight and then adds all the results.

[0130] In the embodiment of the present application, the time attention calculation module obtains the association weight of each candidate machine group operation entity from step 404. At the same time, the value vector generated in advance for each candidate machine group operation entity in step 401 is obtained. Then, the module multiplies the association weight of each candidate machine group operation entity with its corresponding value vector, and adds all the weighted value vectors of the candidate machine group operation entities to obtain a comprehensive context vector.

[0131] Step 406: Taking the module length of the weighted sum result as the association strength value between the threat entity and the corresponding machine group operation entity.

[0132] In step 406, the module length is the length of a vector, which represents the size of the vector in mathematics.

[0133] In the embodiment of the present application, the time attention calculation module calculates the module length of the comprehensive context vector obtained in step 405. The calculated module length value is finally defined as the association strength value between the current threat entity and all candidate machine group operation entities appearing within the fixed time window.

[0134] The following is a specific example:

[0135] The key calculation details of the time attention calculation module to obtain a higher association strength value are as follows: the system first converts the entity identifier of the threat entity low-level wind shear into a query vector in the form of a numerical list, and converts the entity identifier of the machine group operation entity executing the go-around procedure into a key vector and a value vector. Then, the module retrieves the machine group operation entity executing the go-around procedure as a candidate within a fixed time window after the time stamp corresponding to the threat entity low-level wind shear. Subsequently, the module calculates the dot product between the query vector of the threat entity low-level wind shear and the key vector of the candidate machine group operation entity executing the go-around procedure, to obtain a similarity score indicating that the initial similarity between the two is high. This similarity score is then input into the nonlinear transformation layer of the time attention calculation module for processing, for example, using the Softmax function for standardization, so that the sum of the weights of all candidate operations is one, thereby obtaining a standardized association weight for the operation of executing the go-around procedure, which is high. Then, the module performs a weighted sum operation on the higher association weight and the value vector corresponding to the candidate machine group operation entity executing the go-around procedure. Finally, the module length of the vector obtained by the weighted sum operation is calculated, and its value is taken as the final association strength value between the threat entity low-level wind shear and the machine group operation entity executing the go-around procedure. The higher association strength value accurately quantifies the strong time sequence response relationship between the machine group operation and the threat in this event, laying a solid calculation foundation for subsequent successful screening of the candidate machine group operation entity and final determination of the management state.

[0136] In the embodiments of the present application, the temporal correlation between threats and operations is converted into quantifiable intensity values through the above-mentioned vectorized representation and attention calculation process. This process can capture complex contextual semantic information, thereby providing reliable and accurate data for subsequent precise screening of effective management actions.

[0137] To further improve the objectivity, credibility and tamper resistance of threat scoring, in some embodiments, step 103: the threat scoring matrix is generated by using a group decision algorithm in combination with distributed scoring nodes under a blockchain consensus mechanism, from three dimensions of training effect, likelihood of occurrence and severity of consequences, to score each type of threat at three levels, including:

[0138] Step 501: create a corresponding number of distributed scoring nodes according to the total number of threat types, and each distributed scoring node is composed of a scoring group consisting of flight operation experts, flight instructors and risk analysts.

[0139] In step 501, the distributed scoring node is an independent logical unit in the network responsible for scoring a specific type of threat. The scoring group is a team composed of personnel with different professional backgrounds who make scoring decisions within the distributed scoring node.

[0140] In the embodiments of the present application, the system first counts the total number of threat types that need to be evaluated. Then, according to this total number, the same number of distributed scoring nodes is created. Each distributed scoring node is assigned a specific type of threat as its scoring object. Inside each node, a scoring group is formed, which must include members from three different professional fields: flight operation experts, flight instructors and risk analysts, to ensure the comprehensiveness of the evaluation perspective.

[0141] Step 502: Each distributed scoring node scores the specified type of threat in three scoring dimensions: training effect, likelihood of occurrence, and severity of consequences, using an improved Delphi method, to generate a node consensus scoring vector.

[0142] In step 502, the improved Delphi method is a method of guiding the expert group to gradually reach a consensus through multiple rounds of anonymous opinion solicitation and feedback, which is a specific type of group decision algorithm. The node consensus scoring vector is an ordered list consisting of three numbers, which represent the consensus scores of the scoring node on the three dimensions of training effect, likelihood of occurrence and severity of consequences for the threat it is responsible for.

[0143] In the embodiments of the present application, within each distributed scoring node, the coordinator initiates the scoring task for the threat type it is assigned to. The panelists first submit initial scores and reasons for the three dimensions independently and anonymously. After the system collects these scores, it calculates the statistical distribution of each dimension and feeds back to all panelists together with the anonymous reasons. The panelists can revise their scores by referring to the feedback information. This process repeats for multiple rounds until the preset convergence condition is met, such as the scoring opinions tend to be consistent or the maximum number of rounds is reached. Finally, the average value of the scores of each dimension in the last round is taken as the consensus score of the node, and is combined into a node consensus score vector in a fixed order.

[0144] Step 503: Submit the node consensus score vector generated by all distributed scoring nodes to the blockchain network, and perform multiple rounds of exchange and verification through the practical Byzantine fault tolerance consensus mechanism. When more than a preset proportion of nodes agree on the score of a specific dimension of a specific threat, the score of the specific dimension is recorded as a consensus score on the distributed ledger.

[0145] In step 503, the blockchain network is a distributed database maintained by multiple independent nodes, and once the data is recorded, it is difficult to tamper with. The practical Byzantine fault tolerance consensus mechanism is an algorithm that can ensure that all honest nodes reach an agreement on the data even if there are a small number of malicious nodes or faulty nodes in the network. The consensus score is the authoritative score of a specific threat in a specific dimension that has been globally verified and finally confirmed by the blockchain network, and is the consensus score of the same dimension by the members of the three roles of flight operation experts, flight instructors and risk analysts. The distributed ledger is an unalterable database in the blockchain network for recording all consensus results.

[0146] In the embodiments of the present application, after completing the internal consensus, all distributed scoring nodes submit the node consensus score vector generated by them to the blockchain network. The nodes in the network perform multiple rounds of communication and verification on these score vectors through the practical Byzantine fault tolerance consensus mechanism. In this process, the nodes exchange and check each other's scores for the same threat and the same dimension. When it is found that more than a preset proportion of nodes in the network give the same or very close scores for a certain dimension of a certain threat, this score value is recognized as a globally valid consensus score and is permanently and unalterably recorded on the distributed ledger of the blockchain.

[0147] Step 504: For each threat type, extract the consensus scores of the threat type in the three dimensions from the distributed ledger of the blockchain, and combine them to form a score vector corresponding to the threat type.

[0148] In step 504, the score vector of a threat type is an ordered list of three numbers, which are the final consensus scores of the threat type in the three dimensions of training effect, occurrence probability and consequence severity, extracted from the distributed ledger of the blockchain.

[0149] In the embodiment of the present application, after the blockchain network completes the record of the consensus scores of all threats in each dimension, the system queries and reads each threat type's consensus scores in the three dimensions of score from the distributed ledger of the blockchain according to the identifier of the threat type. Then, the system combines the three consensus scores in the order of training effect, occurrence probability and consequence severity to form a score vector representing the final evaluation result of the threat type.

[0150] Step 505: Arrange the score vectors of all threat types in order of threat type identifier to construct a threat score matrix.

[0151] In the embodiment of the present application, the system collects the score vectors generated for all threat types in step 504. Then, the system arranges these score vectors in order according to the natural order of threat type identifiers or a specific sorting rule. Finally, these ordered score vectors together form a complete threat score matrix, which is a key input for subsequent calculation of the comprehensive risk value. After constructing the threat score matrix, the embodiment of the present application can store the Merkle root hash value of the threat score matrix in the block header of the blockchain to ensure its tamper resistance.

[0152] The following is a specific example:

[0153] After the system identifies multiple threat entities such as low-level wind shear communication interruption and completes the management state annotation, a group consensus scoring process is started to generate a threat scoring matrix. Specifically, the system first creates a corresponding number of distributed scoring nodes according to the total number of threat types identified. The distributed scoring node for the low-level wind shear threat group includes a scoring group composed of three types of role members, namely flight operation experts, flight instructors, and risk analysts. The group performs multiple rounds of anonymous independent scoring and reason explanation on the low-level wind shear by using the improved Delphi method, and continuously corrects the opinions by referring to the statistical feedback of each round of scoring. Finally, the group reaches internal consensus in three dimensions of training effect probability and severity of consequences, and generates a node consensus scoring vector, which can be expressed as: training effect dimension score, occurrence probability dimension score, and severity dimension score. Subsequently, all distributed scoring nodes including the nodes processing threats such as low-level wind shear communication interruption submit their generated node consensus scoring vectors to the blockchain network. The nodes in the network perform multiple rounds of exchange and verification through the practical Byzantine fault tolerance consensus mechanism. When more than a preset proportion of nodes agree on the scoring of a specific threat such as low-level wind shear in each dimension, these scores are recorded as consensus scores in the distributed ledger of the blockchain. The system then extracts the consensus scores of each threat type in the three dimensions from the distributed ledger, and combines them to form the corresponding scoring vector of the low-level wind shear. Finally, the system arranges all threat type scoring vectors in the order of threat type identifiers, thereby constructing a threat scoring matrix for subsequent comprehensive risk value calculation. The row data corresponding to the low-level wind shear in the matrix is the final consensus score of the low-level wind shear in the three dimensions.

[0154] In the embodiments of the present application, by combining distributed node organization, improved group decision-making method, and blockchain consensus technology, a decentralized, transparent process, and tamper-proof threat scoring system is constructed. The system effectively gathers the wisdom of experts from multiple fields, ensuring the fairness, authority, and reliability of the scoring results, and laying a solid foundation for subsequent accurate risk assessment.

[0155] To further improve the efficiency and quality of consensus reached within the distributed scoring node, in some embodiments, step 502: scoring of the specified type of threat in the three dimensions of training effect, occurrence probability, and severity of consequences is performed by using the improved Delphi method to generate a node consensus scoring vector, including:

[0156] Step 601: Obtain the scoring task for the specified type of threat initiated by the scoring group coordinator in the distributed scoring node, which contains three dimensions and multiple iteration rules of the improved Delphi method. The multiple iteration rules include an improved convergence condition.

[0157] In step 601, the scoring group coordinator is the person responsible for initiating and managing the scoring process within the distributed scoring node. The scoring task refers to the work order that requires the scoring group to evaluate the specified threat type from the three dimensions of training effectiveness, occurrence probability, and severity of resulting consequences. The multi-round iteration rule defines the number of rounds that the scoring process needs to repeat and the operation specification of each round. The improved convergence condition is a set of more flexible criteria than the traditional method for determining whether the scoring process can be stopped.

[0158] In the embodiments of the present application, within the distributed scoring node, the scoring group coordinator first creates a new scoring task. The task clearly indicates the threat type to be evaluated and specifies the three dimensions on which the evaluation is based. At the same time, the task also details the process of the improved Delphi method to be used, including how to conduct multiple rounds of scoring and feedback, and the improved convergence condition for determining when the scoring can be stopped. After the system obtains and confirms the task, it initiates the subsequent scoring process.

[0159] Step 602: During the execution of the scoring task, the original anonymous scores and original anonymous score reasons for the three dimensions independently submitted by each role member of the scoring group based on the local knowledge base and historical evaluation data in the first round are obtained.

[0160] In step 602, the local knowledge base is a database stored locally by the scoring member, containing his professional knowledge and experience. The historical evaluation data is the scoring data and results accumulated in past similar evaluation tasks. The original anonymous score is the initial score value submitted anonymously by each role member in the first round for the three dimensions. The original anonymous score reason is the explanation written by the member for his initial score.

[0161] In the embodiments of the present application, after the scoring task is initiated, the system sends a scoring invitation to all role members of the scoring group. Based on their own local professional knowledge base and past historical evaluation data, each member independently judges the specified threat on the three dimensions of training effectiveness, occurrence probability, and severity of resulting consequences, forming their own initial scores. At the same time, they also need to write reasons to support their scores. All these initial scores and reasons are submitted to the system in an anonymous manner to ensure the independence of opinions and avoid being influenced by others.

[0162] Step 603: Based on the original anonymous scores and original anonymous score reasons for each dimension, the statistical distribution values of the scores for each dimension are calculated, including the interquartile range, and the statistical distribution values and all role members' original anonymous score reasons are collectively used as feedback information and distributed to all role members for reference by the role members to generate new anonymous scores and new score reasons for each dimension.

[0163] In step 603, the statistical distribution value is a numerical value used to describe the distribution characteristics of a set of data, and the interquartile range is one of them, which reflects the dispersion degree of the middle 50% of the data. The feedback information is the data package distributed to each member after the system integration, which contains the statistical distribution of the score of each dimension and the anonymous scoring reasons of all members.

[0164] In the embodiments of the present application, after the system collects all the original anonymous scores submitted by the members in the first round, the statistical distribution values of all scores of each dimension are calculated, and in particular, the interquartile range is calculated to understand the concentration and dispersion of the scores. Then, the system integrates these statistical distribution values with the original anonymous scoring reasons submitted by all members (in an anonymous state) to form a complete feedback information. This feedback information is then distributed to each member of the scoring group for their reference when performing the next round of scoring.

[0165] Step 604: Based on the new scores and new scoring reasons of each dimension, the statistical distribution values of the scores of each dimension are recalculated, and the new scores are revised anonymous scores or original anonymous scores, and the new scoring reasons are revised anonymous scoring reasons or original anonymous scoring reasons.

[0166] In step 604, the new anonymous score is the score that the member may modify and submit anonymously again after referring to the feedback information. The new scoring reason is the explanation written by the member for the modified score, or the original reason is maintained.

[0167] In the embodiments of the present application, after receiving the feedback information, the members of the scoring group refer to the statistical data and the reasons of others in the feedback information to re-examine their previous scores. They can decide to modify their scores and explain new reasons, or they can stick to the original scores and reasons. The members anonymously submit these new scores and reasons to the system again. After the system collects all the new scores in this round, it recalculates the new statistical distribution values of each dimension, including the new interquartile range, as in step 603.

[0168] Step 605: Repeat the distribution operation, the acquisition operation and the recalculation operation until one of the following improved convergence conditions is met: the interquartile range of the scores of each dimension is less than or equal to the preset convergence threshold. The maximum number of iterations is reached. More than a preset proportion of role members indicate that they no longer modify the scores. After meeting the improved convergence condition, the arithmetic mean of the scores of each dimension in the last round is taken as the consensus score of the corresponding distributed scoring node for the threat type, and the consensus scores of all dimensions are combined in the fixed order of training effect, occurrence probability and severity to generate the node consensus score vector.

[0169] In step 605, the improved convergence condition is a flexible standard for judging whether the iteration can be terminated. The dimension score can be understood as an anonymous score.

[0170] In the embodiments of the present application, the system repeatedly performs steps 603 and 604, i.e., distributes the feedback information, obtains new scores and new reasons, and recalculates the statistical distribution values. This loop process will continue until at least one improved convergence condition is met: for example, the interquartile ranges of all dimension scores are small to an acceptable extent, indicating that the opinions have been sufficiently concentrated; or the maximum number of loop rounds set by the system is reached; or a sufficiently high proportion of members explicitly express that they no longer modify their scores. Once any of the convergence conditions is met, the iteration stops. The system takes the arithmetic mean of all member scores in each dimension in the last round as the final consensus score of the node on the threat in this dimension. Finally, the consensus scores of the three dimensions are combined into a vector according to the fixed order of training effect, occurrence probability, and severity, i.e., to generate the node consensus score vector of the node.

[0171] The following is a specific example:

[0172] The scoring group coordinator first initiates a scoring task for low-altitude wind shear in the node, clearly needs to evaluate from the three dimensions of training effect, occurrence probability, and severity of resulting consequences, and informs the multi-round iteration rules and improved convergence conditions. In the first round of scoring, members of three types of roles, flight operation experts, flight instructors, and risk analysts, based on their respective local knowledge bases and historical evaluation data, independently and anonymously submit original anonymous scores and corresponding original anonymous score reasons for the three dimensions. The system then calculates the statistical distribution values of the scores in each dimension, including the interquartile range, based on these original anonymous scores, and integrates these statistical distribution values and all anonymous reasons into feedback information distributed to each member. After referring to this feedback information, some members revise their opinions and generate new anonymous scores and new score reasons for each dimension, and some maintain the original scores and reasons. The system recalculates the statistical distribution values based on the new scores submitted in the new round. This process of distributing feedback, obtaining new scores and new reasons, and recalculating statistical distribution values is repeated for several rounds until the improved convergence condition is met, i.e., the interquartile ranges of the scores in each dimension are less than or equal to the preset convergence threshold, indicating that the expert opinions have been sufficiently concentrated. Then, the system takes the arithmetic mean of all scores in each dimension in the last round as the consensus score of the node on the low-altitude wind shear threat in this dimension, and strictly combines the consensus scores of the three dimensions according to the fixed order of training effect, occurrence probability, and severity to generate the node consensus score vector of the node, which is the basis for subsequent submission to the blockchain network for global consensus.

[0173] In the embodiments of the present application, by introducing the improved Delphi method and setting flexible convergence conditions, the distributed scoring nodes can efficiently and orderly reach a high-quality group consensus on the basis of sufficient communication and information transparency, thereby ensuring the scientificity and representativeness of the node consensus scoring vector.

[0174] To further improve the pertinence and efficiency of the score correction, in some embodiments, step 603: the reference feedback information generates new anonymous scores and new score reasons of each dimension, including:

[0175] Step 701: compare the anonymous scores of the role members in the last round of three dimensions with the statistical distribution values of the last round in the feedback information, to identify whether there is a target dimension based on the comparison result, the target dimension including: the dimension whose anonymous score is outside the consensus interval and the dimension whose quartile distance is greater than a preset divergence threshold.

[0176] In step 701, the consensus interval is a numerical range calculated according to the statistical distribution values of all member scores in the last round, which generally represents the region where the opinions of most members are concentrated. The preset divergence threshold is a critical value preset for judging whether the scoring divergence is too large. The target dimension refers to the specific dimension that needs to be paid special attention to by the scoring member and considered for modification of the score.

[0177] In the embodiments of the present application, the system first calculates a consensus interval for each scoring dimension according to all the anonymous scores collected in the last round. Then, the system compares the anonymous scores of each role member in the last round for each dimension with the consensus interval corresponding to the dimension. At the same time, the system also checks whether the quartile distance of each dimension score exceeds the preset divergence threshold. Through this comparison and checking process, the system identifies two types of dimensions that need to be paid special attention to for each member: one type is the dimension whose score of the member falls outside the consensus interval, and the other type is the dimension whose overall opinion divergence is still large, i.e., the quartile distance exceeds the preset divergence threshold. These identified dimensions are the target dimensions of the current round for the member.

[0178] Step 702: when there is a target dimension, perform score adaptive correction based on the original anonymous score reasons of all role members in the reference feedback information, to obtain new anonymous scores and new score reasons of each dimension.

[0179] In step 702, score adaptive correction refers to the process in which the scoring member adjusts his score independently according to the feedback information, especially the reasons of other members.

[0180] In the embodiments of the present application, when the system identifies one or more target dimensions for a member, the member focuses on reading the content about these target dimensions in the feedback information, especially the original anonymous rating reasons submitted by all other role members (anonymous). The member carefully analyzes the rationality and persuasiveness of these reasons, and reflects whether the previous rating is biased or not considered. Based on this analysis and reflection, the member decides whether to modify the rating on the target dimension and how to modify it. If the rating is modified, a new rating reason needs to be written to explain the reason for the modification; if it is decided not to modify, the original rating reason can be selected to be reiterated or slightly adjusted to strengthen the position. Finally, the member forms a new round of anonymous rating and rating reason for all dimensions (including target dimensions and non-target dimensions) and submits them.

[0181] The following is a specific example:

[0182] The system first compares the expert's last round of anonymous rating on the three dimensions of training effect occurrence possibility severity with the corresponding last round statistical distribution value in the feedback information. The comparison result identifies that the training effect dimension is the target dimension, because the expert's last round of anonymous rating on the training effect is outside the consensus interval determined by the statistical distribution value, and at the same time the quartile range of this dimension rating is greater than the preset divergence threshold, indicating that the experts have a large difference of opinion on this dimension. When this target dimension exists, the expert will make adaptive correction based on the original anonymous rating reasons of all role members contained in the feedback information. He focuses on reading the rating reasons of other anonymous experts on the training effect dimension, which generally emphasize the key role of targeted training in improving the decision-making and manipulation quality of the wind shear modification. After carefully considering these reasons, the expert believes that he may have underestimated the value of training, so he modifies the rating of the training effect dimension towards the consensus interval, generates a new anonymous rating, and elaborates in the new rating reason that his viewpoint change is based on the persuasiveness of the peer reasons and his new understanding of the value of training. As for the occurrence possibility and severity, which are non-target dimensions, he believes that his original rating and reason are reasonable, so he maintains them, and finally gets his new anonymous rating and new rating reason on each dimension in this round.

[0183] In the embodiments of the present application, by accurately identifying the dimensions that need to be focused on and providing targeted reference information, the rating member is guided to reflect and correct efficiently, thereby accelerating the formation process of consensus and improving the scientificity and rationality of the final node consensus rating vector.

[0184] Figure 3 The structure diagram of a training priority determination system based on statistical analysis of aviation event data provided in the embodiments of the present application, and the specific implementation part describes:

[0185] The extraction module 31 is configured to extract aviation unsafe event data from a plurality of heterogeneous data sources.

[0186] The determination module 32 is configured to extract threats, errors and undesired aircraft states corresponding to each aviation unsafe event from the aviation unsafe event data, and determine a management state of the crew to the threats, errors and undesired aircraft states, the management state being a managed state or an unmanaged state.

[0187] The scoring module 33 is configured to perform three-level scoring on each type of threat from three dimensions of training effect, occurrence probability and severity of resulting consequences by using a group decision algorithm in combination with distributed scoring nodes under a blockchain consensus mechanism, and generate a threat scoring matrix.

[0188] The calculation module 34 is configured to count a crew management rate and an actual occurrence frequency of each type of threat, and calculate a comprehensive risk value of each type of threat in combination with the threat scoring matrix, the crew management rate and the actual occurrence frequency of the threat.

[0189] The determination module 35 is configured to sort all threats according to the comprehensive risk value, determine a training priority based on a sorting result, and embed at least one threat with a high training priority and an error associated with the threat into a scene design of a flight training subject.

[0190] The system of the embodiments of the present application is used to implement the foregoing method, and therefore the specific embodiments of the training priority determination system based on aviation event data statistical analysis can be seen from the foregoing embodiment part of the training priority determination method based on aviation event data statistical analysis, and the specific embodiments can be referred to the description of the corresponding embodiment part, which will not be described herein.

[0191] The present application also provides an electronic device, which comprises a memory for storing a computer program and a processor for implementing the steps of the training priority determination method based on aviation event data statistical analysis according to any one of the foregoing embodiments when the computer program is executed.

[0192] The present application also provides a computer readable storage medium, which stores a computer program, and the steps of the training priority determination method based on aviation event data statistical analysis are implemented when the computer program is executed by a processor.

[0193] In an exemplary embodiment, the computer readable storage medium described above can include but is not limited to a U disk, a read-only memory, a random access memory, a mobile hard disk, a magnetic disk or an optical disk and various media that can store computer programs.

[0194] The embodiment of the application further provides a computer program product, the computer program product comprising a computer program, the computer program being executed by a processor to implement the steps in any of the training priority determination method embodiments based on aviation event data statistical analysis.

[0195] Further, the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in the above description in general terms. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0196] The above provides a training priority determination method, system, electronic device and storage medium based on aviation event data statistical analysis. The principles and implementation modes of the application are described herein by applying specific examples. The above description of the embodiments is only to help understand the method and its core idea. It should be pointed out that for those skilled in the art, without departing from the principles of the application, some improvements and modifications can be made to the application, and these improvements and modifications also fall within the protection scope of the application.

Claims

1. A training priority determination method based on statistical analysis of aviation event data, characterized in that, The method comprises the following steps: extracting aviation unsafe event data from multiple heterogeneous data sources; extracting threats, errors and undesired aircraft states corresponding to each aviation unsafe event from the aviation unsafe event data, and determining the management state of the crew to the threats, errors and undesired aircraft states, which is a managed state or an unmanaged state; using a group decision algorithm based on distributed scoring nodes, under a blockchain consensus mechanism, to score each type of threat from three dimensions of training effect, occurrence probability and severity of resulting consequences, and generate a threat scoring matrix; statistically analyzing the crew management rate and actual occurrence frequency of each type of threat, combining the threat scoring matrix, the crew management rate and the actual occurrence frequency of the threat, and calculating the comprehensive risk value of each type of threat; According to the comprehensive risk value, all threats are sorted, and the training priority is determined based on the sorting result, and at least one threat with high training priority and the error associated with the threat is embedded into the scene design of the flight training subject; extracting threats, errors and undesired aircraft states corresponding to each aviation unsafe event from the aviation unsafe event data, and determining the management state of the crew to the threats, errors and undesired aircraft states, comprising: inputting the aviation unsafe event data into a dynamic knowledge graph model constructed based on the threat and error management (TEM) classification method for entity recognition to extract the threats, errors and undesired aircraft states corresponding to each aviation unsafe event; determine the management state of the crew to the threats, errors and undesired aircraft states through a time series attention mechanism, which is a managed state or an unmanaged state; The determination of the management state of the crew to the threats, errors and undesired aircraft states through the time series attention mechanism comprises: Taking the threats, errors and undesired aircraft states corresponding to all aviation unsafe events as threat entities, error entities and undesired aircraft state entities respectively, constructing an event entity sequence in chronological order based on the entity identifiers corresponding to each entity; inputting the event entity sequence into a time series analysis model based on a pre-configured time series attention mechanism, calculating the correlation strength value between each threat entity and all crew operation entities appearing in a preset time period through a time series attention calculation module in the time series analysis model, wherein the crew operation entities include standard procedure response actions and pilot autonomous intervention actions; Through a dynamic threshold screening module in the time series analysis model, dynamically adjusting the screening threshold according to the type attribute of the threat entity to obtain an adjusted threshold, and screening candidate crew operation entities from all crew operation entities whose correlation strength value with each threat entity is greater than the adjusted threshold; Through the action effectiveness verification module in the time series analysis model, the compliance and utility of each candidate crew operation entity are evaluated to determine whether the candidate crew operation entity is an effective management action for the corresponding threat, and an effectiveness determination result is obtained. The state labeling logic module in the time series analysis model labels a corresponding management state for each threat entity based on the validity determination result, and derives and labels the same management state for the associated error entity and the undesired aircraft state entity based on the management state of the threat entity and the causal relationship between the threat entity and the corresponding error entity and undesired aircraft state entity queried from the dynamic knowledge graph.

2. The training priority determination method based on statistical analysis of aviation event data according to claim 1, characterized in that, The time series attention calculation module in the time series analysis model calculates the association strength value between each threat entity in the event entity sequence and all crew operation entities appearing within a preset time period, including: Converting each entity identifier in the event entity sequence into a numerical vector, wherein the vector corresponding to the threat entity is defined as a query vector, and the vector corresponding to the crew operation entity is defined as a key vector; For each threat entity, all appearing crew operation entities within a fixed time window after the corresponding timestamp of the threat entity are retrieved; The dot product between the query vector of the threat entity and the key vector of each candidate crew operation entity is calculated to obtain a similarity score; The similarity score is input into the nonlinear transformation layer of the time series attention calculation module for processing to obtain a standardized association weight; The standardized association weight corresponding to each candidate crew operation entity is weighted and summed with the value vector corresponding to the candidate crew operation entity, wherein the value vector is obtained by linear transformation of the vector corresponding to the candidate crew operation entity through the time series attention calculation module; The length of the weighted sum result is taken as the association strength value between the threat entity and the corresponding crew operation entity.

3. The training priority determination method based on statistical analysis of aviation event data according to claim 1, characterized in that, The threat scoring matrix is generated by using a group decision algorithm based on the training effect, occurrence probability, and severity of the resulting consequences of each type of threat under the blockchain consensus mechanism, including: According to the total number of threat types, a corresponding number of distributed scoring nodes are created, each of which is composed of a scoring group consisting of flight operation experts, flight instructors, and risk analysts; Each distributed scoring node scores the specified type of threat in the three scoring dimensions of training effect, occurrence probability, and severity of resulting consequences by an improved Delphi method, generating a node consensus score vector; The node consensus score vectors generated by all distributed scoring nodes are submitted to the blockchain network, and the practical Byzantine fault tolerance consensus mechanism is used for multiple rounds of exchange and verification. When more than a preset proportion of nodes agree on the specific dimension score of a specific threat, the specific dimension score is recorded as a consensus score to the distributed ledger; For each threat type, the consensus scores of the threat type in the three dimensions are extracted from the blockchain distributed ledger to form a scoring vector for the corresponding threat type; All threat type scoring vectors are arranged in order of threat type identifiers to construct a threat scoring matrix.

4. The training priority determination method based on statistical analysis of aviation event data according to claim 3, characterized in that, The three dimensions of the specified type of threat are scored by the improved Delphi method, a node consensus score vector is generated, including: Obtaining a scoring task for a specified type of threat initiated by a scoring group coordinator at the distributed scoring node, the scoring task containing three dimensions and the multi-round iteration rules of the improved Delphi method, the multi-round iteration rules containing an improved convergence condition; During the execution of the scoring task, the raw anonymous scores and raw anonymous score reasons for the three dimensions independently submitted by each role member of the scoring group based on the local knowledge base and historical evaluation data in the first round are obtained; Based on the raw anonymous scores and raw anonymous score reasons of each dimension, the statistical distribution value of each dimension score is calculated, including the interquartile range, and the statistical distribution value and all role members' raw anonymous score reasons are collectively used as feedback information and distributed to all role members for reference by role members to generate new anonymous scores and new score reasons for each dimension; Based on the new scores and new score reasons of each dimension, the statistical distribution value of each dimension score is recalculated, and the new score is a revised anonymous score or an original anonymous score, and the new score reason is a revised anonymous score reason or an original anonymous score reason; The distribution operation, the acquisition operation and the recalculation operation are repeated until one of the following improved convergence conditions is met: the interquartile range of each dimension score is less than or equal to the preset convergence threshold; the preset maximum iteration round is reached; more than a preset proportion of role members indicate that they no longer modify the score; After meeting the improved convergence condition, the arithmetic mean of the last round of each dimension score is taken as the consensus score of the corresponding distributed scoring node for the threat type, and the consensus scores of all dimensions are combined in the fixed order of training effect, occurrence possibility and severity to generate the node consensus score vector.

5. The training priority determination method based on statistical analysis of aviation event data according to claim 4, characterized in that, The reference feedback information generates new anonymous scores and new score reasons for each dimension, including: Comparing the anonymous scores of the role members in the last three dimensions with the statistical distribution value in the last round of feedback information to identify whether there is a target dimension based on the comparison result, the target dimension including: the dimension whose anonymous score is outside the consensus interval and the dimension whose interquartile range is greater than a preset divergence threshold; When there is a target dimension, the original anonymous score reasons of all role members in the reference feedback information are used for adaptive correction to obtain new anonymous scores and new score reasons for each dimension.

6. A training priority determination system based on statistical analysis of aviation event data, characterized by, It includes: An extraction module is used to extract aviation safety event data from a plurality of heterogeneous data sources; A determination module is used to extract threats, errors and undesirable aircraft states corresponding to each aviation safety event from the aviation safety event data, and determine the management state of the crew to the threats, errors and undesirable aircraft states, the management state being a managed state or an unmanaged state; The scoring module is configured to score each type of threat from three dimensions of training effect, occurrence probability and severity of resulting consequences by using a group decision algorithm in combination with distributed scoring nodes under a blockchain consensus mechanism, and generate a threat scoring matrix; The calculating module is configured to calculate a comprehensive risk value of each type of threat by combining the threat scoring matrix, the unit management rate and the actual occurrence frequency of the threat; The determining module is configured to sort all threats according to the comprehensive risk value, determine a training priority based on the sorting result, and embed at least one threat with a high training priority and an error associated with the threat into a scene design of a flight training subject; The threats, errors and undesired aircraft states corresponding to each aviation safety incident are extracted from the aviation safety incident data, and the management state of the crew to the threats, errors and undesired aircraft states is determined, including: The aviation safety incident data is input into a dynamic knowledge graph model based on the TEM classification method to extract the threats, errors and undesired aircraft states corresponding to each aviation safety incident; The management state of the crew to the threats, errors and undesired aircraft states is determined by a time sequence attention mechanism, and the management state is a managed state or an unmanaged state; The determination of the management state of the crew to the threats, errors and undesired aircraft states by the time sequence attention mechanism includes: The threats, errors and undesired aircraft states corresponding to all aviation safety incidents are respectively taken as threat entities, error entities and undesired aircraft state entities, and event entity sequences are constructed in chronological order based on entity identifiers corresponding to each entity; The event entity sequence is input into a time series analysis model based on a pre-configured time sequence attention mechanism, and the correlation strength value between each threat entity in the event entity sequence and all crew operation entities appearing in a preset time period is calculated by a time sequence attention calculation module in the time series analysis model, wherein the crew operation entities include standard procedure response actions and pilot autonomous intervention actions; A dynamic threshold screening module in the time series analysis model dynamically adjusts a screening threshold according to the type attribute of the threat entity, obtains an adjusted threshold, and screens candidate crew operation entities from all crew operation entities, wherein the correlation strength value of each threat entity is greater than the adjusted threshold; An action effectiveness verification module in the time series analysis model performs compliance and utility evaluation on each candidate crew operation entity to determine whether the candidate crew operation entity is an effective management action for the corresponding threat, and obtains an effectiveness determination result. The state labeling logic module in the time series analysis model labels a corresponding management state for each threat entity based on the effectiveness determination result, and derives and labels the same management state for the associated error entity and the undesired aircraft state entity based on the management state of the threat entity and the causal association relationship between the threat entity and the corresponding error entity and undesired aircraft state entity queried from the dynamic knowledge graph.

7. An electronic device, comprising: Comprise: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the training priority determination method based on aviation event data statistical analysis according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the training priority determination method based on aviation event data statistical analysis according to any one of claims 1 to 5.

Images