General aviation flight training risk assessment method, system and device and storage medium

Abstract

The invention discloses a general aviation flight training risk assessment method, system and device and a storage medium. The method comprises the following steps: acquiring survey report information of a flight training unsafe event, and dividing the survey report information into a model corpus and an instance corpus; forming a model corpus containing labels; constructing a risk identification model based on a deep learning algorithm; constructing a flight training risk evolution analysis network; calculating node parameters of accident cause factors according to a flight training risk evolution analysis network; and sorting the risk nodes of the flight training unsafe events by integrating the node parameters, thereby completing the steps of risk assessment of general aviation flight training and the like. The method has the advantages that the accident cause theory, deep learning and the complex network form a hybrid model to be applied to general aviation flight training risk assessment, subjectivity can be reduced, influence factors of flight training unsafe events can be effectively assessed, and research results can provide scientific basis for formulating systems such as general aviation flight training safety management.

Description

Technical Field

[0001] The present invention belongs to the technical field of general aviation accident risks, and in particular relates to a general aviation flight training risk assessment method, system, device and storage medium. Background Art

[0002] As an important part of civil aviation, the safety level of general aviation is directly related to the stability and development of the entire civil aviation industry. However, due to the multifaceted interaction of different personnel, the diversity of organizational structures, the need to adapt to technological progress, and the gradual allocation of resources brought about by the continuous development of the low-altitude economy, the general aviation field faces many risks and challenges, resulting in the possibility of unsafe incidents, which is particularly evident in flight training.

[0003] At present, there are few studies on the factors affecting the risk of general aviation flight training. At the same time, the widely used complex network risk assessment methods lack objectivity in the selection of indicators and network structure, making it difficult to provide effective support for flight safety risk management. Therefore, the application of data-driven complex network modeling methods in the study of flight training accident risks has important practical significance. Summary of the invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a general aviation flight training risk assessment method, system, device and storage medium, which can more accurately reflect the comprehensive status of accident-causing factors as nodes in a complex general aviation network, and provide effective support for flight safety risk management.

[0005] In order to achieve the above-mentioned purpose, the general aviation flight training risk assessment method provided by the present invention performs the following steps in sequence:

[0006] 1) Obtain investigation report information on unsafe flight training incidents and divide it into a model corpus and an instance corpus;

[0007] 2) Label the risk factors and risk relationships in the above model corpus according to the accident causation theory to form a model corpus containing labels;

[0008] 3) Construct a risk identification model based on a deep learning algorithm, and then input the above-mentioned model corpus containing labels into the risk identification model for training and evaluation to obtain a trained risk identification model;

[0009] 4) Input the instance corpus into the trained risk identification model, extract the risk factors and risk relationships in the investigation report information, and construct a flight training risk evolution analysis network based on the extracted risk factors and risk relationships;

[0010] 5) Calculating the node parameters of the accident causal factors according to the flight training risk evolution analysis network; the node parameters include node degree value, closeness centrality and betweenness centrality;

[0011] 6) Based on the above node parameters, the risk nodes of flight training unsafe events are ranked, thereby completing the risk assessment of general aviation flight training.

[0012] In step 1), the method of obtaining the investigation report information of the flight training unsafe incident and dividing it into a model corpus and an instance corpus is:

[0013] Obtaining the investigation report information of the unsafe flight training incident from a preset website; wherein the investigation report information includes the cause, analysis and type of the unsafe flight training incident;

[0014] The types of unsafe flight training incidents include:

[0015] Loss of control in flight (LOC-I), controlled flight into terrain (CFIT), abnormal runway contact, system / component failure and others;

[0016] Among them, the risks brought by loss of control during flight include casualties, aircraft damage and social impact, and its severity far exceeds that of other types of accidents;

[0017] Accordingly, the above investigation report information is divided into two parts according to the types of unsafe flight training incidents:

[0018] The investigation report information of non-in-flight loss of control unsafe events in flight training is used as the model corpus to train and evaluate the risk identification model. The investigation report information of in-flight loss of control unsafe events in flight training is used as the instance corpus to extract the risk factors and risk relationships, which are used to verify the actual effect of the risk identification model.

[0019] In step 2), the risk factors and risk relationships in the above model corpus are labeled according to the accident causation theory, and the method for constructing the model corpus containing the labels is:

[0020] Label the risk factors in the above model corpus according to the traditional PMEO classification method; label the risk relationships in the model corpus according to the behavioral safety 2-4 model;

[0021] The risk factors include:

[0022] Human Factors (PER) - covers a wide range of issues related to flight personnel, including pilots, flight instructors, maintenance personnel, and all other personnel directly concerned with flight safety;

[0023] Aircraft Factors (AIR) - refers to problems related to the aircraft itself, including defects or deficiencies in the aircraft's flight conditions, design, manufacture and maintenance;

[0024] Environmental factors (ENV) – involve challenges related to the flight environment, including weather conditions, geographic location, air traffic conditions and other external environmental factors;

[0025] Management factors (ORG) – challenges related to the operations of the airline or related organization, including deficiencies or inadequacies in management systems, human resource policies and operational processes;

[0026] The risk relationship is labeled as "Cause" according to the safety 2-4 model chain structure, thereby constructing a model corpus containing labels.

[0027] In step 3), the risk identification model based on the deep learning algorithm is constructed, and then the model corpus containing the labels is input into the risk identification model for training and evaluation. The method for obtaining the trained risk identification model is:

[0028] The risk identification model includes an enhanced BERT module, a bidirectional long short-term memory (Bi-LSTM) module, and a conditional random field (CRF) module;

[0029] The enhanced BERT introduces the adversarial training algorithm FreeLB and uses the attention distribution of all words in the above-mentioned model corpus containing labels to perform direct training in the parameter space to enhance the generalization ability of the enhanced BERT module when processing unbalanced accident text data and complete the encoding of the word vector sequence;

[0030] The bidirectional long short-term memory (Bi-LSTM) module includes forward and backward long short-term memory (LSTM) layers, which processes the above word vector sequence bidirectionally to form a comprehensive feature representation to capture the context at both ends of the word vector sequence;

[0031] The conditional random field (CRF) module processes the feature representation generated by the bidirectional long short-term memory module through state transition, transition probability and label decoding, and uses a cross entropy loss function to measure the difference between the label sequence predicted by the risk identification model and the actual label sequence;

[0032] The training process adopts the Early Stopping strategy, including dividing the data set, monitoring indicators, setting stop conditions and rolling back model parameters;

[0033] The data set division is to further divide the model corpus into a training data set, a validation data set and a test data set;

[0034] The monitoring indicator is to use the cross entropy loss function and the fitting evaluation index to monitor the performance of the risk identification model on the validation data set;

[0035] The calculation formula of the cross entropy loss function is:

[0036]

[0037] in, represents the cross entropy loss, y represents the true label probability value, Represents the label probability value predicted by the risk identification model;

[0038] The fitting evaluation indicators include Precision, Recall and F1_score, and the calculation formula is:

[0039]

[0040] Among them, correct_entities represents the number of labels correctly predicted by the model, pred_entities represents the number of labels predicted by the model, and true_entities represents the number of true labels;

[0041] The setting of the stopping condition is to set a threshold, which indicates the number of rounds at which the model performance is allowed to no longer improve on the validation dataset. If the model performance does not improve for multiple consecutive iterations on the validation dataset, the training will stop.

[0042] The rollback model parameters are usually rolled back to the state in which the model has the best performance on the validation dataset when the training stops, that is, rolled back to the model parameters corresponding to the iteration with the best performance.

[0043] In step 4), the method of inputting the instance corpus into the trained risk identification model, extracting risk factors and risk relationships in the investigation report information, and constructing a flight training risk evolution analysis network based on the extracted risk factors and risk relationships is:

[0044] Taking the risk factors of unsafe events in flight training as nodes and risk relationships as connecting lines, nodes are created in Gephi software, and the tools in Gephi software are used to draw a complex network to obtain the structure of the flight training risk evolution analysis network.

[0045] In step 5), the method for calculating the node parameters of the accident causal factors according to the flight training risk evolution analysis network is:

[0046] The node parameters include node degree, closeness centrality and betweenness centrality;

[0047] The node degree value represents the total number of edges connected to the node, including the node's in-degree and the node's out-degree; wherein the node's in-degree represents the number of incoming edges, and the out-degree represents the number of outgoing edges;

[0048] The calculation formula of the closeness centrality is:

[0049]

[0050] Among them, CC i represents closeness centrality, d i represents the average value of the shortest path distance from node i to all other nodes in the network; d ij represents the shortest path length between node i and node j in the flight training risk evolution analysis network, N i represents the set of nodes directly connected to node i in the flight training risk evolution analysis network;

[0051] The calculation formula of the betweenness centrality is:

[0052]

[0053] Among them, BC i represents the betweenness centrality, g st represents the sum of the shortest paths between node s and node t in the flight training risk evolution analysis network, It represents the number of times the shortest path between node s and node t in the flight training risk evolution analysis network passes through node i.

[0054] In step 6), the method of sorting the risk nodes of flight training unsafe events by comprehensively analyzing the above node parameters, thereby completing the risk assessment of general aviation flight training is:

[0055] The VIKOR algorithm (EW-VIKOR) fused with the entropy weight method is used to comprehensively evaluate the impact of each node parameter on risk evolution analysis, including building a decision matrix, standardizing the decision matrix, calculating entropy values ​​and weights, determining ideal solutions and negative ideal solutions, determining utility metrics and regret metrics, and calculating ranking indexes:

[0056] Based on the three node parameters of node degree, closeness centrality and betweenness centrality of each node mentioned above, the decision matrix is ​​constructed as follows:

[0057]

[0058] Where D represents the decision matrix, a m (b k ) represents the value of the mth node parameter under the kth parameter index;

[0059] In order to eliminate the dimensional differences between node parameters, the above decision matrix is ​​standardized, and the formula is as follows:

[0060]

[0061] Among them, x mk It represents the value of the mth node parameter under the kth parameter index after standardization; n represents the total number of nodes.

[0062] According to the value x of the mth node parameter under the kth parameter index after standardization mk Calculate the entropy value using the following formula:

[0063]

[0064] Among them, E k Represents the information entropy of the kth parameter index;

[0065] Indicates the proportion of the mth node parameter under the kth parameter index;

[0066] According to the information entropy E of the kth parameter index k Calculate the weight, the formula is as follows:

[0067]

[0068] Among them, W k represents the weight of the kth parameter index; l represents the total number of parameter indicators.

[0069] According to the value x of the mth node parameter under the kth parameter index after standardization mk Calculate the ideal solution and negative ideal solution, the formula is as follows:

[0070]

[0071] in, represents an ideal solution; represents a negative ideal solution;

[0072] B j =x mk -x kmin represents the benefit function; x kmin Indicates the minimum value of the kth parameter index;

[0073] C j =x kmax -x mk represents the cost function; x kmax Indicates the maximum value of the kth parameter index;

[0074] According to the weight W of the kth parameter index k 、Ideal solution and negative ideal solution The utility metric and regret metric are calculated as follows:

[0075]

[0076] Among them, S m represents the utility measure; R m represents the regret metric;

[0077] The ranking index is calculated based on the utility metric and the regret metric, and the formula is as follows:

[0078] Q m =v(S m -S * ) / (S - -S * )+(1-v)(R m -R * ) / (R - -R * )

[0079] Among them, Q m Indicates the ranking index of the mth node parameter; S * =min m S m ; S - =max m S m ; R * =min m R m ; R - =max m R m ; v is 0.5;

[0080] Finally, the importance of risk nodes is ranked according to the above ranking index, and the risk factors are divided into three levels: high, medium and low using the threshold segmentation method based on the median M. Nodes exceeding 1.1 times the median M are defined as high-risk nodes, nodes below 0.9 times the median M are defined as low-risk nodes, and nodes between high-risk nodes and low-risk nodes are defined as medium-risk nodes, thereby completing the risk assessment of general aviation flight training.

[0081] The general aviation flight training risk assessment system provided by the present invention comprises:

[0082] Data collection module, used to obtain investigation report information on unsafe flight training incidents;

[0083] The risk identification module is used to extract risk factors and risk relationships from the investigation report information using deep learning natural language processing technology;

[0084] A network construction module is used to construct a flight training risk evolution analysis network based on risk factors and risk relationships;

[0085] The risk assessment module is used to calculate the node parameters of accident causation factors according to the flight training risk evolution network, and to evaluate the risk nodes of unsafe flight training events based on the comprehensive node parameters.

[0086] The general aviation flight training risk assessment device provided by the present invention comprises:

[0087] at least one processor;

[0088] at least one memory for storing at least one program;

[0089] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0090] The computer-readable storage medium provided by the present invention stores a program executable by a processor, and the program executable by the processor is used to execute the above method when executed by the processor.

[0091] The beneficial effects of the present invention are: applying the hybrid model composed of accident causation theory, deep learning and complex networks to the risk assessment of general aviation flight training can reduce subjectivity and effectively evaluate the influencing factors of unsafe flight training incidents. The research results can provide a scientific basis for the formulation of general aviation flight training safety management and other systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the embodiments of the present invention or the drawings of related technical solutions in the prior art are introduced below. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0093] Figure 1 A flow chart of a method for risk assessment of unsafe events in general aviation flight training provided by an embodiment of the present invention;

[0094] Figure 2 A performance evaluation result diagram of the general aviation flight training unsafe event risk identification algorithm provided by an embodiment of the present invention; wherein (a) is the verification loss, and (b) is the fitting accuracy;

[0095] Figure 3 A schematic diagram of a risk evolution network model for unsafe events in general aviation flight training provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0096] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0097] like Figure 1 As shown, the general aviation flight training risk assessment method provided in this embodiment includes the following steps performed in sequence:

[0098] 1) Obtain investigation report information on unsafe flight training incidents and divide it into a model corpus and an instance corpus;

[0099] This embodiment obtains the investigation report information of flight training unsafe events from the official website of the National Transportation Safety Board of the United States (https: / / www.ntsb.gov / ), with a total of 1748 records (retrieval time: January 1, 2012 to December 31, 2021), including the possible causes, analysis and types of flight training unsafe events;

[0100] The types of unsafe flight training incidents include:

[0101] Loss of control in flight (LOC-I), controlled flight into terrain (CFIT), abnormal runway contact, system / component failure and others;

[0102] Among them, the risks brought by loss of control during flight include casualties, aircraft damage and social impact, and its severity far exceeds that of other types of accidents;

[0103] Based on this, the above investigation report information is divided into two parts according to the types of unsafe flight training incidents:

[0104] The investigation report information of 1430 unsafe flight training events without loss of control during flight was used as the model corpus to train and evaluate the risk identification model. The investigation report information of 318 unsafe flight training events without loss of control during flight was used as the instance corpus to extract the risk factors and risk relationships, which was used to verify the actual effect of the risk identification model.

[0105] 2) Label the risk factors and risk relationships in the above model corpus according to the accident causation theory to form a model corpus containing labels;

[0106] The risk factors in the model corpus are annotated according to the traditional PMEO classification method, that is, the investigation report information is analyzed by human factors (PER), aircraft factors (AIR), environmental factors (ENV) and management factors (ORG);

[0107] The risk factors include:

[0108] Human Factors (PER) - covers a wide range of issues related to flight personnel, including pilots, flight instructors, maintenance personnel, and all other personnel directly concerned with flight safety;

[0109] Aircraft Factors (AIR) - refers to problems related to the aircraft itself, including defects or deficiencies in the aircraft's flight conditions, design, manufacture and maintenance;

[0110] Environmental factors (ENV) – involve challenges related to the flight environment, including weather conditions, geographic location, air traffic conditions and other external environmental factors;

[0111] Management factors (ORG) – challenges related to the operations of the airline or related organization, including deficiencies or inadequacies in management systems, human resource policies and operational processes;

[0112] According to the chain structure of the safety 2-4 model, the risk relationship in the model corpus is labeled as "Cause", thereby constructing a model corpus containing labels.

[0113] 3) Construct a risk identification model based on a deep learning algorithm, and then input the above-mentioned model corpus containing labels into the risk identification model for training and evaluation to obtain a trained risk identification model;

[0114] The risk identification model includes an enhanced BERT module, a bidirectional long short-term memory (Bi-LSTM) module, and a conditional random field (CRF) module;

[0115] The enhanced BERT introduces the adversarial training algorithm FreeLB and uses the attention distribution of all words in the above-mentioned model corpus containing labels to perform direct training in the parameter space to enhance the generalization ability of the enhanced BERT module when processing unbalanced accident text data and complete the encoding of the word vector sequence;

[0116] The bidirectional long short-term memory (Bi-LSTM) module includes forward and backward long short-term memory (LSTM) layers, which processes the above word vector sequence bidirectionally to form a comprehensive feature representation to capture the context at both ends of the word vector sequence;

[0117] The conditional random field (CRF) module processes the feature representation generated by the bidirectional long short-term memory module through state transition, transition probability and label decoding, and uses a cross entropy loss function to measure the difference between the label sequence predicted by the risk identification model and the actual label sequence;

[0118] The training process adopts the Early Stopping strategy, including dividing the data set, monitoring indicators, setting stop conditions and rolling back model parameters;

[0119] The data set division is to further divide the model corpus into a training data set, a validation data set and a test data set;

[0120] The monitoring indicator is to use the cross entropy loss function and the fitting evaluation index to monitor the performance of the risk identification model on the validation data set; the performance evaluation results are as follows Figure 2 As shown, the Val_loss (i.e., validation loss) of the risk identification model reaches the lowest point, indicating that no overfitting occurs, and the risk identification model also achieves the best fitting accuracy, Precision = 95.26%, Recall = 93.69%, and F1 score = 94.47%.

[0121] 4) Input the instance corpus into the trained risk identification model, extract the risk factors and risk relationships in the investigation report information, and construct a flight training risk evolution analysis network based on the extracted risk factors and risk relationships;

[0122] 4.1) Identification of risk factors

[0123] Among the risk factors affecting general aviation flight training, personnel-related factors include: improper aircraft control (P1), improper use of equipment (P2), improper flight planning (P3), improper pre-flight inspection (P4), lack of communication (P5), violation of procedures (P6), improper remedial measures (P7), misjudgment (P8), distraction (P9), failure to monitor the flight (P10), lack of situational awareness (P11), spatial disorientation (P12), visual illusion (P13), self-pressure (P14), lack of flight experience (P15), lack of aircraft knowledge (P16), flight fatigue (P17), and maintenance errors (P18);

[0124] Among the risk factors affecting general aviation flight training, aircraft-related factors include: engine failure (A1), fuel system failure (A2), landing gear system failure (A3), flight control system failure (A4), electrical system failure (A5), communication system failure (A6), aircraft structure failure (A7), aircraft rotor failure (A8), loss of directional control (A9), exceeding the critical angle of attack (A10), failure to maintain airspeed (A11), improper landing flaps (A12), failure to maintain runway heading (A13), and aircraft design defects (A14);

[0125] Among the risk factors affecting general aviation flight training, those related to the environment include: collision with obstacles (E1), collision with birds (E2), runway surface conditions (E3), insufficient light (E4), low visibility (E5), turbulence (E6), wind (E7), convective weather (E8), and damage to airport facilities (E9);

[0126] Among the risk factors affecting general aviation flight training, those related to management include: lack of airport regulations (O1), lack of safety culture (O2), inadequate training procedures (O3), poor supervision (O4), and inadequate safety management (O5);

[0127] 4.2) Identification of risk relationships

[0128] Taking the risk information “improper aircraft control (P1)” as an example, the risk information corresponding to the risk relationship “Cause” includes: loss of directional control (A9), exceeding the critical angle of attack (A10), failure to maintain airspeed (A11), incorrect landing flaps (A12), failure to maintain runway heading (A13), collision with obstacles (E1), and collision with birds (E2).

[0129] 4.3) Construction of flight training risk evolution analysis network

[0130] like Figure 3 As shown in the figure, the risk factors of unsafe flight training events are used as nodes, and the risk relationships are used as connecting lines. Nodes are created in the Gephi software, and the tools in the Gephi software are used to draw a complex network to obtain the structure of the flight training risk evolution analysis network.

[0131] 5) Calculate the node parameters of accident causal factors based on the flight training risk evolution analysis network;

[0132] The node parameters include node degree, closeness centrality and betweenness centrality;

[0133] According to the topological structure of the flight training risk evolution analysis network, the node degree values ​​of each risk factor are shown in Table 1:

[0134] Table 1

[0135]

[0136]

[0137] From Table 1, we can see that the overall degree value of personnel risk affecting flight training risk is relatively high, indicating that personnel risk is at the core of general aviation flight training. At the same time, the out-degree value of personnel risk is greater than its in-degree value, indicating that the relationship between personnel risk and other risk factors is mainly the impact of personnel risk information on other risk factors; the in-degree value of aircraft risk is higher than the out-degree value, which indicates that aircraft risk is more susceptible to the impact of other risk points; the out-degree value and out-degree value of environmental risk are roughly equal, indicating that environmental risk has a significant impact on the occurrence and consequences of unsafe events in the risk chain;

[0138] According to the topological structure of the flight training risk evolution analysis network, the closeness centrality of each risk factor is shown in Table 2:

[0139] Table 2

[0140]

[0141]

[0142] From Table 2, we can see that the risk factors such as “distraction” (P9), “poor lighting conditions” (E4), “failure to monitor the flight” (P10), “inappropriate corrective action” (P7), and “flight control system failure” (A4) have high closeness centrality, indicating that they are close to the center of the network, that is, they are closely related to other risk factors;

[0143] According to the topological structure of the flight training risk evolution analysis network, the closeness centrality of each risk factor is shown in Table 3:

[0144]

[0145]

[0146] From Table 3, we can see that “procedural violation” (P6), “poor judgment” (P8), “communication system failure” (A6), “improper aircraft control” (P1), and “incorrect landing flaps” (A12) have high betweenness centrality. These risk factors appear as key mediators of risk information diffusion, indicating that they have a significant impact on the diffusion dynamics in flight training.

[0147] 6) Based on the above node parameters, the risk nodes of flight training unsafe events are ranked, thereby completing the risk assessment of general aviation flight training.

[0148] This embodiment uses the VIKOR algorithm (EW-VIKOR) integrated with the entropy weight method to comprehensively evaluate the impact of various node parameters on risk evolution analysis, including constructing a decision matrix, standardizing the decision matrix, calculating entropy values ​​and weights, determining ideal solutions and negative ideal solutions, determining utility metrics and regret metrics, and calculating ranking indexes; finally, the risk nodes are ranked in importance according to the above ranking indexes, and the risk factors are divided into three levels: high, medium, and low using a threshold segmentation method based on the median M. Nodes with a value exceeding 1.1 times the median M are defined as high-risk nodes, nodes with a value below 0.9 times the median M are defined as low-risk nodes, and nodes between high-risk nodes and low-risk nodes are defined as medium-risk nodes.

[0149] Taking high-risk factors as an example, the ranking of their comprehensive measurement values ​​is shown in Table 4.

[0150] Table 4

[0151] node Comprehensive Metrics P1 Improper aircraft control 4.72 A9 lost control of direction 4.63 A10 exceeds critical angle of attack 4.43 A11 failed to maintain airspeed 4.36 A13 failed to maintain runway heading 4.26 E2 collides with a bird 4.17 P6 Violation of procedures 4.11 A12 incorrect landing flaps 4.10 E1 collides with an obstacle 4.08

[0152] General aviation flight training risk prevention suggestions:

[0153] For high-risk factors that have the potential to lead to serious consequences, the focus should be on prevention and the strategic allocation of resources to reduce the likelihood and severity of these incidents. This approach requires a comprehensive strategy that includes risk assessment, early warning systems and the implementation of strong safety measures. In addition, regular monitoring and auditing of high-risk areas is essential to ensure the effectiveness of these preventive measures;

[0154] For medium risk factors, a balance must be struck between the expense and benefits of risk control measures. It is essential to regularly monitor these factors and flexibly adjust control strategies based on changing circumstances. In addition, strengthening education and training for relevant personnel is also essential to improve awareness and response capabilities.

[0155] For low risk factors, it is best to adopt simplified control measures to reduce resource expenditure. The management of such risks can be incorporated into daily operations and maintenance activities. Although the impact of these risks is relatively small, you should continue to be vigilant and look for opportunities to improve control measures.

[0156] The general aviation flight training risk assessment system provided in this embodiment includes:

[0157] Data collection module, used to obtain investigation report information on unsafe flight training incidents;

[0158] The risk identification module is used to extract risk factors and risk relationships from the investigation report information using deep learning natural language processing technology;

[0159] A network construction module is used to construct a flight training risk evolution analysis network based on risk factors and risk relationships;

[0160] The risk assessment module is used to calculate the node parameters of accident causation factors according to the flight training risk evolution network, and to evaluate the risk nodes of unsafe flight training events based on the comprehensive node parameters.

[0161] The general aviation flight training risk assessment device provided in this embodiment includes:

[0162] at least one processor;

[0163] at least one memory for storing at least one program;

[0164] When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

[0165] The computer-readable storage medium provided in this embodiment stores a program executable by a processor, and the program executable by the processor is used to perform the above method when executed by the processor.

[0166] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above-mentioned implementation mode. Technical personnel familiar with the field can also make various equivalent deformations or substitutions without violating the spirit of the present invention. These equivalent deformations or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A general aviation flight training risk assessment method, characterized by: The general aviation flight training risk assessment method comprises the following steps performed in sequence: 1) Obtain investigation report information on unsafe flight training incidents and divide it into a model corpus and an instance corpus; 2) Label the risk factors and risk relationships in the above model corpus according to the accident causation theory to form a model corpus containing labels; 3) Construct a risk identification model based on a deep learning algorithm, and then input the above-mentioned model corpus containing labels into the risk identification model for training and evaluation to obtain a trained risk identification model; 4) Input the instance corpus into the trained risk identification model, extract the risk factors and risk relationships in the investigation report information, and construct a flight training risk evolution analysis network based on the extracted risk factors and risk relationships; 5) Calculate the node parameters of accident causal factors based on the flight training risk evolution analysis network; 6) Based on the above node parameters, the risk nodes of flight training unsafe events are ranked, thereby completing the risk assessment of general aviation flight training.

2. The general aviation flight training risk assessment method according to claim 1, characterized in that: In step 1), the method of obtaining the investigation report information of the flight training unsafe incident and dividing it into a model corpus and an instance corpus is: Obtaining the investigation report information of the unsafe flight training incident from a preset website; wherein the investigation report information includes the cause, analysis and type of the unsafe flight training incident; The types of unsafe flight training incidents include: Loss of control in flight, controlled flight into terrain, abnormal runway contact, system / component failure and others; The above investigation report information is divided into two parts according to the type of unsafe flight training incidents: The investigation report information of non-in-flight loss of control unsafe events in flight training is used as the model corpus to train and evaluate the risk identification model. The investigation report information of in-flight loss of control unsafe events in flight training is used as the instance corpus to extract the risk factors and risk relationships, which are used to verify the actual effect of the risk identification model.

3. The general aviation flight training risk assessment method according to claim 1, characterized in that: In step 2), the risk factors and risk relationships in the above model corpus are labeled according to the accident causation theory, and the method for constructing the model corpus containing the labels is: Label the risk factors in the above model corpus according to the traditional man-machine-environment-management classification method; Label the risk relationships in the model corpus according to the Behavioral Safety 2-4 model; The risk factors include: Human Factors - covers a wide range of issues related to flight personnel, including pilots, flight instructors, maintenance personnel, and all other personnel directly concerned with flight safety; Aircraft factors - refers to problems related to the aircraft itself, including defects or deficiencies in the aircraft's flight conditions, design, manufacture and maintenance; Environmental factors – involve challenges related to the flight environment, including weather conditions, geographic location, air traffic conditions and other external environmental factors; Management factors – challenges related to the operations of the airline or related organization, including deficiencies or inadequacies in management systems, human resource policies and operational processes; The risk relationship is labeled as "Cause" according to the safety 2-4 model chain structure, thereby constructing a model corpus containing labels.

4. The general aviation flight training risk assessment method according to claim 1, characterized in that: In step 3), the risk identification model based on the deep learning algorithm is constructed, and then the model corpus containing the labels is input into the risk identification model for training and evaluation. The method for obtaining the trained risk identification model is: The risk identification model includes an enhanced BERT module, a bidirectional long short-term memory module, and a conditional random field module; The enhanced BERT module introduces the adversarial training algorithm FreeLB and uses the attention distribution of all words in the above-mentioned model corpus containing labels to perform direct training in the parameter space to enhance the generalization ability of the enhanced BERT module when processing unbalanced accident text data and complete the encoding of the word vector sequence; The bidirectional long short-term memory module includes forward and backward long short-term memory layers, which processes the above word vector sequence bidirectionally to form a comprehensive feature representation to capture the context at both ends of the word vector sequence; The conditional random field module processes the feature representation generated by the bidirectional long short-term memory module through state transition, transition probability and label decoding, and uses a cross entropy loss function to measure the difference between the label sequence predicted by the risk identification model and the actual label sequence; The training process adopts the Early Stopping strategy, including dividing the data set, monitoring indicators, setting stop conditions and rolling back model parameters; The data set division is to further divide the model corpus into a training data set, a validation data set and a test data set; The monitoring indicator is to use the cross entropy loss function and the fitting evaluation index to monitor the performance of the risk identification model on the validation data set; The calculation formula of the cross entropy loss function is: in, represents the cross entropy loss, y represents the true label probability value, Represents the label probability value predicted by the risk identification model; The fitting evaluation indicators include Precision, Recall and F1_score, and the calculation formula is: Among them, correct_entities represents the number of labels correctly predicted by the model, pred_entities represents the number of labels predicted by the model, and true_entities represents the number of true labels; The setting of the stopping condition is to set a threshold, which indicates the number of rounds at which the model performance is allowed to no longer improve on the validation dataset. If the model performance does not improve for multiple consecutive iterations on the validation dataset, the training will stop. The rollback model parameters are usually rolled back to the state in which the model has the best performance on the validation data set when the training stops, that is, rolled back to the model parameters corresponding to the iteration with the best performance.

5. The general aviation flight training risk assessment method according to claim 1, characterized in that: In step 4), the method of inputting the instance corpus into the trained risk identification model, extracting risk factors and risk relationships in the investigation report information, and constructing a flight training risk evolution analysis network based on the extracted risk factors and risk relationships is: Taking the risk factors of unsafe events in flight training as nodes and risk relationships as connecting lines, nodes are created in Gephi software, and the tools in Gephi software are used to draw a complex network to obtain the structure of the flight training risk evolution analysis network.

6. The general aviation flight training risk assessment method according to claim 1, characterized in that: In step 5), the method for calculating the node parameters of the accident causal factors according to the flight training risk evolution analysis network is: The node parameters include node degree, closeness centrality and betweenness centrality; The node degree value represents the total number of edges connected to the node, including the node's in-degree and the node's out-degree; wherein the node's in-degree represents the number of incoming edges, and the out-degree represents the number of outgoing edges; The calculation formula of the closeness centrality is: Among them, CC i represents closeness centrality, d i represents the average value of the shortest path distance from node i to all other nodes in the network; d ij represents the shortest path length between node i and node j in the flight training risk evolution analysis network, N i represents the set of nodes directly connected to node i in the flight training risk evolution analysis network; The calculation formula of the betweenness centrality is: Among them, BC i represents the betweenness centrality, g st represents the sum of the shortest paths between node s and node t in the flight training risk evolution analysis network, It represents the number of times the shortest path between node s and node t in the flight training risk evolution analysis network passes through node i.

7. The general aviation flight training risk assessment method according to claim 1, characterized in that: In step 6), the method of sorting the risk nodes of flight training unsafe events by comprehensively analyzing the above node parameters, thereby completing the risk assessment of general aviation flight training is: The VIKOR algorithm integrated with the entropy weight method is used to comprehensively evaluate the impact of each node parameter on risk evolution analysis, including building a decision matrix, standardizing the decision matrix, calculating entropy values ​​and weights, determining ideal solutions and negative ideal solutions, determining utility metrics and regret metrics, and calculating ranking indexes: Based on the three node parameters of node degree, closeness centrality and betweenness centrality of each node mentioned above, the decision matrix is ​​constructed as follows: Where D represents the decision matrix, a m (b k ) represents the value of the mth node parameter under the kth parameter index; In order to eliminate the dimensional differences between node parameters, the above decision matrix is ​​standardized, and the formula is as follows: Among them, x mk It represents the value of the mth node parameter under the kth parameter index after standardization; n represents the total number of nodes. According to the value x of the mth node parameter under the kth parameter index after standardization mk Calculate the entropy value using the following formula: Among them, E k Represents the information entropy of the kth parameter index; Indicates the proportion of the mth node parameter under the kth parameter index; According to the information entropy E of the kth parameter index k Calculate the weight, the formula is as follows: Among them, W k represents the weight of the kth parameter index; l represents the total number of parameter indexes; According to the value x of the mth node parameter under the kth parameter index after standardization mk Calculate the ideal solution and negative ideal solution, the formula is as follows: in, represents an ideal solution; represents a negative ideal solution; B j =x mk -x kmin represents the benefit function; x kmin Indicates the minimum value of the kth parameter index; C j =x kmax -x mk represents the cost function; x kmax Indicates the maximum value of the kth parameter index; According to the weight W of the kth parameter index k 、Ideal solution and negative ideal solution The utility metric and regret metric are calculated as follows: Among them, S m represents the utility measure; R m represents the regret metric; The ranking index is calculated based on the utility metric and the regret metric, and the formula is as follows: Q m =v(S m -S * ) / (S - -S * )+(1-v)(R m -R * ) / (R - -R * ) Among them, Q m Indicates the ranking index of the mth node parameter; S * =min m S m ; S = max m S m ; R * =min m R m ; R = max m R m ; v is 0.5; Finally, the importance of risk nodes is ranked according to the above ranking index, and the risk factors are divided into three levels: high, medium and low using the threshold segmentation method based on the median M. Nodes exceeding 1.1 times the median M are defined as high-risk nodes, nodes below 0.9 times the median M are defined as low-risk nodes, and nodes between high-risk nodes and low-risk nodes are defined as medium-risk nodes, thereby completing the risk assessment of general aviation flight training.

8. A general aviation flight training risk assessment system, characterized by: The general aviation flight training risk assessment system includes: Data collection module, used to obtain investigation report information on unsafe flight training incidents; The risk identification module is used to extract risk factors and risk relationships from the investigation report information using deep learning natural language processing technology; A network construction module is used to construct a flight training risk evolution analysis network based on risk factors and risk relationships; The risk assessment module is used to calculate the node parameters of accident causation factors according to the flight training risk evolution network, and to evaluate the risk nodes of unsafe flight training events based on the comprehensive node parameters.

9. A general aviation flight training risk assessment device, characterized in that: The general aviation flight training risk assessment device comprises: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to perform the method described in any one of claims 1 to 7 when executed by the processor.

Images