Deep learning-based flight over-limit event record text classification method and system

Abstract

The application discloses a kind of based on deep learning's flight over-limit event record text classification method and system, by integrating " obtain historical over-limit associated collaborative data list, construct over-limit associated data network, input aircraft over-limit prediction model and predict, with the prediction result as pre-over-limit event record text classification mark " full-process technical features, formed " data acquisition-network construction-feature extraction-prediction classification " complete closed loop, effectively solved the core defects that traditional method exists in background technology, over-limit event inducing factor is difficult to accurately locate, pre-over-limit event record text classification is disjointed with over-limit prediction, prediction accuracy is insufficient, responsibility is fuzzy, pre-over-limit text classification is not targeted and early warning lag, overall realizes the accurate classification mark of pre-over-limit event record text and the accurate prediction of flight over-limit event, simultaneously improves the reuse value and management efficiency of pre-over-limit event record text.

Description

Technical Field

[0001] This invention relates to the field of civil aviation flight safety technology, specifically to a deep learning-based method and system for classifying flight over-limit event records, used for risk prediction and prevention of over-limit events during aircraft operation, providing technical support for flight safety prevention and control. Background Technology

[0002] Flight exceeding limits (such as exceeding altitude, speed, and attitude limits) is a core hidden danger affecting civil aviation flight safety. Its occurrence is not due to operational errors by a single position, but is closely related to the quality of collaborative work among various positions (flight crew, air traffic control, dispatch, etc.) throughout the entire aircraft operation process. The efficiency of collaboration between positions, the accuracy of instruction exchange, and the smoothness of responsibility handover directly determine the probability of exceeding limits incidents. Therefore, positions and their collaborative work have a decisive impact on exceeding limits. Currently, in the field of civil aviation flight safety management, no one has systematically considered the core issue of "the correlation between positional collaboration and exceeding limits incidents," nor has any relevant technical solution been proposed to address this problem, resulting in significant shortcomings in the prevention and management of exceeding limits incidents.

[0003] Traditional methods have three major flaws in their management models when over-limit incidents occur: First, the determination of responsibility is unclear, making it impossible to identify which positions and their collaborative behaviors are related to the incident; second, responsibility tracking is slow, requiring a significant amount of time to review the operational records of each position, making it difficult to quickly identify the core responsible party; and third, the determination of responsibility is flawed, ignoring the collaborative nature of over-limit incidents and often attributing responsibility to a single position. In reality, many over-limit incidents are the result of improper coordination among multiple positions, accumulating over a long period before finally erupting. Such one-sided determination of responsibility cannot fundamentally prevent similar over-limit incidents from recurring.

[0004] Meanwhile, existing text classification methods related to over-limit events all employ traditional text classification techniques. However, over-limit event records involved in civil aviation flight procedures are characterized by large data volumes, complex structures, and numerous textual features. Traditional text classification methods struggle to efficiently process such data, and their classification accuracy and efficiency fail to meet actual management needs. More importantly, for ongoing flight events, complete flight record data has not yet been collected. Traditional text classification methods, relying on the characteristics of complete text data, cannot promptly mark over-limit events in the flight record, making it difficult to provide early warnings of over-limit risks. This application addresses these gaps and deficiencies by mining historical collaborative and operational data from various positions to predict the over-limit situation of current flight events and pre-label the flight records (over-limit records) of these events. This increases management personnel's attention to current flights and improves the targeting of management efforts, allowing for early avoidance of over-limit risks and significantly enhancing flight safety. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method and system for classifying flight over-limit event record texts, in order to solve the technical problems in the prior art, such as low accuracy in predicting flight over-limit events, lack of specificity in classifying pre-over-limit event record texts, inability to achieve early prevention and control of over-limit risks, and inability to push classification labels in real time. This invention achieves accurate classification and labeling of pre-over-limit event record texts, accurate prediction of over-limit events, and real-time push of classification labels, thereby improving the level of civil aviation flight safety management and the efficiency of pre-over-limit event record text control.

[0006] In a first aspect, embodiments of the present invention provide a deep learning-based method for classifying flight over-limit event record texts, including: Obtain a list of historical over-limit association and collaboration data for all relevant positions involved in the current flight's operation; the list of historical over-limit association and collaboration data includes position terminal device ID information, position name, position personnel information, collaboration data, flight over-limit association annotations, and over-limit event occurrence tags; the collaboration data includes information interaction records between the position's terminal device and other position terminal devices, records of executed instructions, information interaction frequency between the position's terminal device and other terminal devices, and execution instruction efficiency; Construct an over-limit association data network based on a historical list of over-limit associated collaborative data; Input the over-limit correlation data network into the pre-trained aircraft over-limit prediction model and output the aircraft over-limit prediction results; The aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives the over-limit associated data network; the prediction layer extracts the node features, edge features, adjacency features, and label association features of the over-limit associated data network; the over-limit associated data network is adjusted based on the node features, edge features, adjacency features, and label association features to obtain the over-limit key network; and the aircraft over-limit prediction result is predicted based on the node features, edge features, adjacency features, label association features, and the over-limit key network. The output layer is used to output the aircraft over-limit prediction results; The aircraft over-limit prediction results are used as the classification and labeling text of the pre-over-limit event records during the aircraft's operation.

[0007] Optionally, the method further includes sending the pre-exceedance event record text with classification and annotation to the management center.

[0008] Optionally, based on the list of out-of-limit collaborative data, an out-of-limit data network is constructed, including: The network of over-limit association data is constructed based on the positions of each flight during the current flight operation. Each node is recorded as the corresponding position name. The network edges are constructed based on the information interaction records in the collaborative data. The weights of the network edges are obtained based on the collaborative data, the flight over-limit association annotations, and the over-limit event occurrence labels.

[0009] Optionally, based on collaborative data, flight over-limit association annotations, and over-limit event occurrence labels, the weights of network edges are obtained, including: The flight over-limit association weight and the label association weight are obtained based on the flight over-limit association annotation and the over-limit event occurrence label, respectively; The information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are normalized respectively. The normalized information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are then weighted and summed to obtain the weight of the network edge.

[0010] Optionally, the excess-limit association data network can be adjusted based on node features, edge features, adjacency features, and label association features to obtain the excess-limit key network, including: The risk assessment score of a node is obtained by adjusting the node features, edge features, adjacency features, and label association features. Based on point features, edge features, adjacency features, and label association features, the degree of influence of network edges is evaluated, and the out-of-limit event labeling for each edge is determined. The out-of-limit event influence labeling includes Class I edges, Class II edges, and Class III edges. An over-critical network is generated based on the importance score of each node and the importance label of each edge.

[0011] Optionally, based on node features, edge features, adjacency features, label association features, and the over-limit key network, the aircraft over-limit prediction result is predicted, including: Node features, edge features, adjacency features, and label association features are aggregated and normalized to form a set of linked features; The core prediction feature set is obtained based on the linked feature set; The core predictive feature set is integrated with the adjacency relationship of the over-limit critical network, and the comprehensive features of the nodes are extracted. Based on the attention fusion mechanism, the final fusion features of nodes are obtained based on the comprehensive features of nodes; The final fused features of the nodes are input into the logistic regression prediction model, which outputs the probability of flight over-limit events corresponding to each piece of collaborative data. The prediction results for aircraft exceeding limits are generated based on the probability of occurrence.

[0012] Secondly, this application also provides a deep learning-based text classification system for flight over-limit event records, including: The acquisition module is used to acquire a list of historical over-limit association and collaboration data for all relevant positions involved in the current flight's operation. The list of historical over-limit association and collaboration data includes the position's terminal device ID information, position name, personnel information, collaboration data, flight over-limit association annotations, and over-limit event occurrence tags. The collaboration data includes information interaction records between the position's terminal device and other position's terminal devices, records of executed instructions, the frequency of information interaction between the position's terminal device and other terminal devices, and the efficiency of executed instructions. The network construction module is used to build an over-limit association data network based on a historical list of over-limit association collaborative data; The prediction module is used to input the over-limit correlation data network into a pre-trained aircraft over-limit prediction model and output the aircraft over-limit prediction result. The aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives the over-limit correlation data network; the prediction layer extracts the node features, edge features, adjacency features, and label association features of the over-limit correlation data network; adjusts the over-limit correlation data network based on the node features, edge features, adjacency features, and label association features to obtain the over-limit key network; predicts the aircraft over-limit prediction result based on the node features, edge features, adjacency features, label association features, and the over-limit key network; and the output layer outputs the aircraft over-limit prediction result. The classification and labeling module is used to classify and label the pre-excess event record text of the aircraft operation process based on the aircraft excess prediction results.

[0013] Optionally, the system further includes: The sending module is used to classify and label the pre-exceedance event record text and send it to the management center.

[0014] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects: This invention also provides a deep learning-based method and system for classifying flight over-limit event records, which obtains a historical over-limit associated collaborative data list of all relevant positions involved in the current flight's operation. The historical over-limit associated collaborative data list includes position terminal device ID information, position name, personnel information, collaborative data, flight over-limit association annotations, and over-limit event occurrence tags. The collaborative data includes information interaction records between the position's terminal device and other position terminal devices, execution command records, the information interaction frequency between the position's terminal device and other terminal devices, and execution command efficiency. Based on the historical over-limit associated collaborative data list, an over-limit associated data network is constructed; the over-limit associated data is then... The network input is a pre-trained aircraft over-limit prediction model, which outputs aircraft over-limit prediction results. The aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives over-limit associated data networks. The prediction layer extracts node features, edge features, adjacency features, and label association features from the over-limit associated data networks. Based on the node features, edge features, adjacency features, and label association features, the over-limit associated data networks are adjusted to obtain the over-limit key network. Based on the node features, edge features, adjacency features, label association features, and the over-limit key network, the aircraft over-limit prediction results are predicted. The output layer outputs the aircraft over-limit prediction results. The aircraft over-limit prediction results are used as the text classification label for the pre-over-limit event record of the aircraft's operation process.

[0015] Based on the above technical solutions, for each specific technical feature: The technical feature of "obtaining a historical list of over-limit associated collaborative data, including the terminal device ID of the position, position information, collaborative data, flight over-limit association annotations, and over-limit event occurrence tags" solves the problem in the background technology that the collaborative data collection is incomplete and the data correlation is poor, which makes it impossible to associate the collaborative behavior of each position with over-limit events. It realizes the standardized integration of over-limit associated data and provides accurate data support for subsequent network construction and prediction.

[0016] By constructing an over-limit associated data network based on a historical over-limit associated collaborative data list, the limitations of the single-position perspective in the background technology are eliminated. With positions as nodes and position interactions as edges, the traditional method cannot effectively associate collaborative data of multiple positions, and can accurately represent the collaborative linkage relationship of each position.

[0017] By employing the technical feature of "inputting the over-limit associated data network into a pre-trained aircraft over-limit prediction model, the model's input layer receives network data, the recognition layer extracts node features, edge features, adjacency features, and label association features and adjusts the network to obtain the over-limit key network, and the output layer outputs the prediction results," this approach solves the problems of existing models in the background technology, such as the inability to effectively integrate multi-dimensional features, weak feature extraction targeting, and low prediction accuracy. It achieves accurate extraction of multi-dimensional features and focus on the over-limit key network, significantly improving the over-limit prediction accuracy.

[0018] By using the technical feature of "using the aircraft over-limit prediction results as the classification and labeling of pre-over-limit event record texts," the shortcomings of the background technology in classifying pre-over-limit event record texts—namely, the lack of specificity in classification, the disconnect between classification and prediction, and the low reusability—are resolved. This achieves a deep integration of pre-over-limit event record text classification and labeling with over-limit prediction, which facilitates subsequent retrieval, review, and management of pre-over-limit texts. Furthermore, it provides managers with clear guidance on over-limit risks through precise classification and labeling, enabling them to avoid over-limit risks in advance. At the same time, it fills the gap in the background technology, which cannot promptly classify and label pre-over-limit texts that have not completed over-limit records. This effectively solves the problems of vague responsibility assignment, delayed early warning, and chaotic management of pre-over-limit texts in traditional methods, significantly improving the level of civil aviation flight safety management and the efficiency of pre-over-limit event record text control.

[0019] The technical solution of this application, as a whole, integrates the technical features of the entire process of "acquiring a list of historical over-limit associated collaborative data, constructing an over-limit associated data network, inputting the data into an aircraft over-limit prediction model for prediction, and using the prediction results as the classification and labeling of pre-over-limit event record texts." This forms a complete closed loop of "data acquisition - network construction - feature extraction - prediction and classification," effectively solving the core defects of traditional methods in the background technology, such as difficulty in accurately locating the inducing factors of over-limit events, disconnect between the classification of pre-over-limit event record texts and over-limit prediction, insufficient prediction accuracy, ambiguous liability determination, lack of specificity in pre-over-limit text classification, and delayed early warning. Overall, it achieves accurate classification and labeling of pre-over-limit event record texts and accurate prediction of flight over-limit events. It not only breaks the limitations of a single position's perspective and accurately represents the collaborative linkage relationship of multiple positions, but also achieves efficient integration of multi-dimensional features and focus on core risks. It fills the gap in traditional methods that cannot timely classify and label pre-over-limit texts that have not completed over-limit records, while improving the reusability and management efficiency of pre-over-limit event record texts, clarifying the collaborative inducing factors of over-limit events, and solving the problem of ambiguous liability determination. Attached Figure Description

[0020] Figure 1 This is a flowchart of a deep learning-based method for classifying flight over-limit event records, provided in an embodiment of the present invention. Detailed Implementation

[0021] In the research on the classification and labeling of flight over-limit incident records and the early warning of flight over-limit incidents, it was found that:

[0022] The primary core issue encountered in text classification methods is the contradiction between the "co-inducing characteristics" of civil aviation flight over-limit events and the "single-faceted and fragmented" processing mode of existing technologies. This can be summarized into three main challenges: First, the inducing factors of over-limit events are difficult to pinpoint accurately. Flight over-limit events are not caused by errors in a single position, but rather by the cumulative result of improper coordination among multiple positions such as flight crew, air traffic control, and dispatch. Existing technologies cannot effectively link the collaborative data of each position with the inherent correlation of over-limit events, leading to ambiguous accountability and delayed early warnings. Second, the classification of collaborative data is disconnected from over-limit prediction. The collaborative data generated by civil aviation flights is large in volume and complex in structure, including various types of information such as job interactions, command execution, and over-limit labeling. Existing methods can only achieve simple data classification and cannot link the classification results with over-limit risks, resulting in low data reuse value. Thirdly, the prediction accuracy is insufficient. When processing multi-job collaborative features, existing deep learning models either cannot effectively integrate multi-dimensional features such as nodes, edges, and adjacencies, or the feature extraction is not targeted enough, making it difficult to focus on the core collaborative features that induce over-limit events. This leads to a large deviation between the prediction results and the actual scenario, which cannot meet the needs of early prevention and control.

[0023] To address the above issues, we first attempted two mature existing technical solutions, but neither achieved the desired results:

[0024] The first approach uses traditional text classification algorithms (such as SVM and Naive Bayes) combined with single-job operation data for over-limit prediction. This approach focuses only on the operation records of a single job, completely ignoring the decisive impact of inter-job collaboration on over-limit events. Furthermore, traditional text classification algorithms cannot efficiently handle the high-dimensionality and strong correlation characteristics of flight collaboration data. In tests, it was found that the accuracy of over-limit prediction was only about 65%, which could not accurately identify over-limit risks caused by improper collaboration among multiple jobs, and could not achieve targeted classification and labeling of data.

[0025] The second approach involves constructing a job-specific collaborative network (GCN) based on the basic graph network model (such as a standard GCN) commonly used in predicting the remaining life of aero-engines. While this approach achieves preliminary correlation of collaborative data across multiple job roles, it fails to integrate flight over-limit association annotations and over-limit event occurrence labels. The network edge weights are calculated based solely on simple interaction frequencies, which cannot accurately characterize the correlation strength between collaborative behavior and over-limit events. Furthermore, the lack of an effective feature fusion mechanism leads to feature redundancy and the obscuring of useful information. In testing, the prediction accuracy only improved to 72%, and it was unable to accurately locate high-risk collaborative nodes and edges. It still cannot solve the core problems of difficulty in assigning responsibility and slow early warning, and it falls far short of the actual needs of civil aviation flight safety management.

[0026] In addition, we also tried using the LSTM model to process collaborative time series data, but this model has the drawbacks of low sequence computation efficiency, difficulty in parallel processing, and inability to effectively capture spatial collaborative relationships between positions. It also cannot meet the needs of real-time prediction and accurate classification of flight over-limit events.

[0027] Based on the shortcomings of the above attempts, we reflected that to solve the core problem, the key lies in breaking the limitations of "single feature, single position, single model" and finding an integrated logic of "collaborative data association - core feature extraction - over-limit prediction - classification labeling". Ultimately, based on the core understanding that "job collaboration is the core cause of over-limit events", we developed a key ingenuity - using "over-limit association" as the core link, we deeply bind multi-dimensional information such as multi-job collaborative data, over-limit labeling, and historical tags to build a network model that can accurately represent collaborative risks. At the same time, through layer-by-layer optimization of "feature selection - fusion - prediction", we achieve deep integration of collaborative data classification and over-limit prediction. Specifically, the core of the ingenuity lies in three points: First, it abandons the "single-position" perspective and constructs a network of out-of-limit related data by using multiple positions as nodes and position interactions as edges. It deeply binds edge weights with out-of-limit association annotations and historical labels, enabling the network to accurately reflect the correlation strength between collaborative behavior and out-of-limit risks. Second, it innovatively adopts an "integrated feature aggregation + core feature screening" model. It first binds and integrates the four major features of nodes, edges, adjacency, and labels, and then screens out core features that are highly related to out-of-limit events, eliminating redundant information and improving prediction targeting. This is different from the shortcomings of existing models that treat all features equally. Third, it introduces a cross-layer graph attention fusion mechanism, which allocates feature weights based on the risk of nodes and the influence of edges, allowing features of high-risk nodes and high-risk edges to receive key attention. At the same time, it adopts simplified graph convolution operations to balance prediction accuracy and computational efficiency, ultimately achieving a closed loop of "network construction - feature extraction - prediction classification". This not only solves the problem of non-targeted classification of collaborative data, but also significantly improves the accuracy of out-of-limit prediction, thus proposing the complete technical solution of this application. This ingenious design not only fits the actual scenario of civil aviation flights, but also makes up for the core deficiencies of existing technologies, enabling early warning of over-limit risks and precise management of over-limit risk data.

[0028] The present invention will now be described in detail with reference to the accompanying drawings.

[0029] Example 1

[0030] like Figure 1 As shown, this embodiment of the invention provides a deep learning-based method for classifying flight over-limit event record texts, the method comprising:

[0031] S101: Obtain a list of historical over-limit associated collaborative data for all relevant positions involved in the operation of the current flight.

[0032] In this embodiment, the historical over-limit association collaboration data list includes job terminal device ID information, job name, job personnel information, collaboration data, flight over-limit association annotations, and over-limit event occurrence tags. Each job corresponds to a set of job terminal device ID information, job name, job personnel information, collaboration data, flight over-limit association annotations, and over-limit event occurrence tags. The collaboration data includes information interaction records between the job's terminal device and the terminal devices of other jobs, records of executed instructions, the frequency of information interaction between the job's terminal device and other terminal devices, and the efficiency of executed instructions.

[0033] Specifically, the Civil Aviation Flight Data Management System collects historical out-of-limit correlation and collaboration data from all relevant positions (including core positions such as flight crew, air traffic control, dispatch, and ground staff) during the current flight's operation. The collection scope includes collaborative data from similar routes and aircraft types within the three months prior to the current flight, ensuring the data's relevance and reference value. After collection, the data undergoes deduplication, noise reduction, and completion preprocessing (for missing execution command efficiency data, the average of similar data from the same position is used for completion; for abnormal information interaction frequency data, the 3σ criterion is used to remove outliers before completion), ultimately forming a standardized list of historical out-of-limit correlation and collaboration data.

[0034] The historical over-limit association collaborative data list specifically includes the job terminal device ID information, job name, job personnel information, collaborative data, flight over-limit association annotations, and over-limit event occurrence tags, as detailed below:

[0035] (1) Terminal device ID information for each position: The unique identifier of the terminal for each position is uniformly assigned by the civil aviation management system. For example, the flight crew terminal ID is FLT-001 and the air traffic control terminal ID is ATC-101, which is used to accurately associate the collaborative data of each position.

[0036] (2) Job title: Specify the specific job title, including flight crew, air traffic control, dispatch, ground staff, etc., to build network nodes.

[0037] (3) Personnel information: including personnel number, name, and job responsibilities. The personnel number is uniformly compiled by the civil aviation management system. The job responsibilities clearly define the core work of the position during flight (such as the flight crew being responsible for executing flight instructions and adjusting flight attitude, and air traffic control being responsible for issuing altitude and speed instructions), which is used to associate the position's operation with the risk of exceeding limits.

[0038] (4) Collaborative data: This includes three types of data, all extracted in real time from the Civil Aviation Flight Data Management System to ensure data authenticity and timeliness, as follows:

[0039] Information interaction records: All information interaction content between a certain terminal and other terminals, including instruction sending time, instruction content, response time, interaction results, etc., such as the altitude adjustment instruction sent by air traffic control to the flight crew and the flight crew's response record.

[0040] Records of executed instructions: Detailed records of each position's execution of instructions from superiors or other positions, including instruction receipt time, execution start time, execution completion time, execution result (qualified / unqualified), deviation value, etc.

[0041] Information interaction frequency and instruction execution efficiency: Information interaction frequency is the number of times information is exchanged between two terminal devices per unit time (hour), which is automatically calculated by the system; Instruction execution efficiency is scored on a 100-point scale and is comprehensively evaluated by managers based on indicators such as instruction execution deviation and execution time. The scoring criteria are as follows: No execution deviation and execution time within the specified range (≤30 seconds) is 90-100 points; Execution deviation ≤5% and execution time 30-60 seconds is 70-89 points; Execution deviation >5% or execution time >60 seconds is below 70 points.

[0042] (5) Flight Exceedance Correlation Labeling: A 100-point scoring system is adopted. Civil aviation safety management personnel combine historical exceedance events to evaluate the degree of correlation between each position and the position's collaborative behavior and the exceedance event. Specifically, it includes the degree of position participation (0-10 points) and the degree of correlation of instruction deviation (0-10 points). The average of the two scores is the flight exceedance correlation labeling score for that position. The higher the degree of correlation, the higher the score.

[0043] (6) Over-limit event label: Use binary labels (0 or 1). 0 indicates that no over-limit event has occurred in the historical scenario corresponding to this collaborative data, and 1 indicates that an over-limit event has occurred in the historical scenario corresponding to this collaborative data. The label is automatically marked by the system based on the historical over-limit records or manually marked.

[0044] All data in this application comes from actual historical and real-time records of civil aircraft operations, requiring no additional external data. The model architecture is concise, the calculation methods for each step are clear and quantifiable, and specific implementation plans are provided for each unrefined step, ensuring that those skilled in the art can fully implement this method according to the specification. It is easy to implement and can quickly achieve standardized classification and labeling of pre-excess event record text. S102: Construct an excess-limit-related data network based on the historical excess-limit-related collaborative data list.

[0045] In this embodiment, each position during the current flight's operation is considered a node in the over-limit association data network. Each node is recorded as the corresponding position name. Network edges are constructed based on information interaction records in the collaborative data. The weights of the network edges are obtained based on the collaborative data, flight over-limit association annotations, and over-limit event occurrence labels. For example:

[0046] Node construction: Each relevant position during the current flight's operation is used as a node in the over-limit association data network. Each node is recorded as the corresponding position name (such as "flight crew" or "air traffic control"). At the same time, the node is associated with the position's terminal device ID information, personnel information, flight over-limit association annotation data, and the frequency of historical over-limit events. The association method is to use the above information (position terminal device ID information, personnel information, flight over-limit association annotation data, and the frequency of historical over-limit events) as attribute parameters of the node and store them in the node's attribute list for easy feature extraction and risk assessment.

[0047] Network edge construction: Traverse all information interaction records in the historical over-limit association collaboration data list. If there are information interaction records between the terminal devices of two positions (regardless of the number of interactions, as long as there are valid interaction records and the interaction content is related to flight operations), then a network edge is constructed between the nodes corresponding to these two positions. If there are no information interaction records between the terminal devices of two positions, or if the interaction records are unrelated to flight operations (such as private information interaction), then no network edge is constructed, ensuring that the existence of the edge is directly related to the collaboration of positions and the risk of over-limit.

[0048] Network edge weight assignment: Flight violation association weights and tag association weights are obtained based on flight violation association annotations and violation event occurrence labels, respectively. Then, the information interaction frequency, execution command efficiency, and tag association weights are normalized. Finally, the normalized information interaction frequency, execution command efficiency, and tag association weights are weighted and summed to obtain the network edge weights. Specifically, the network edge weights are calculated using a method of "multi-dimensional data fusion + normalization + weighted summation."

[0049] First, calculate the flight over-limit association weight and the label association weight separately:

[0050] Flight violation association weight calculation: Based on flight violation association annotation data, this weight is used to characterize the inherent association strength between collaborative edges and violation events. The specific calculation formula is: Flight violation association weight = (Sum of flight violation association annotation scores of the nodes at both ends of the edge) / 2 × 10, with the result rounded to one decimal place. For example, if the annotation score of the flight crew node is 8.5 and the annotation score of the air traffic control node is 8.5, then the flight violation association weight of the network edge between them is (8.5 + 8.5) / 2 × 10 = 85.0.

[0051] Label association weight calculation: Calculated based on the label data of over-limit events, it is used to characterize the probability of a collaborative edge historically inducing an over-limit event. The specific calculation formula is: Label association weight = (Number of collaborative data with label 1 in the data corresponding to this edge / Total number of collaborative data corresponding to this edge) × 100. The calculation result is rounded to one decimal place. If the total number of collaborative data corresponding to this edge is 0, the label association weight is set to 0. For example, if a network edge corresponds to 10 collaborative data, of which 3 have label 1, then the label association weight = (3 / 10) × 100 = 30.0 points.

[0052] Then, normalization processing is performed: the information interaction frequency and execution command efficiency in the collaborative data, as well as the flight over-limit association weight and tag association weight calculated in the first step, are normalized. The min-max normalization method is used to uniformly adjust the value range of all parameters to 0-100 points to eliminate the difference in parameter magnitude. The specific calculation formula is: normalized value = (original value - minimum value of the parameter) / (maximum value of the parameter - minimum value of the parameter) × 100. The calculation result is rounded to one decimal place.

[0053] Specific operation: First, calculate the maximum and minimum values ​​of information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight corresponding to all network edges. For example, the maximum value of information interaction frequency is 10 times / hour and the minimum value is 1 time / hour. If the information interaction frequency of a certain edge is 6 times / hour, then its normalized value is (6-1) / (10-1)×100≈55.6 points.

[0054] Then, a weighted summation is performed: the normalized information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are weighted and summed to obtain the final weights of the network edges, specifically including:

[0055] Determine the weighting coefficients: Through 5-fold cross-validation calibration and combined with the needs of over-limit prediction, determine the weighting coefficients of each parameter, specifically: information interaction frequency 0.15, execution command efficiency 0.25, flight over-limit association weight 0.25, and tag association weight 0.35 (the tag association weight has the highest proportion, highlighting the reference value of historical over-limit events).

[0056] Calculate the final weight: Final weight = (normalized information interaction frequency × 0.15) + (normalized execution instruction efficiency × 0.25) + (normalized flight over-limit association weight × 0.25) + (normalized label association weight × 0.35). The calculation result is rounded to two decimal places. The larger the weight value, the higher the degree of abnormality of the collaborative edge and the stronger the association with the over-limit event.

[0057] By constructing an over-limit association data network, integrating historical over-limit association collaborative data, flight over-limit association annotations, and over-limit event occurrence labels, and optimizing edge weight calculation; by adjusting the network to obtain the over-limit key network, focusing on high-risk collaborative features; and innovatively adopting a combination logic of "integrated feature aggregation + core feature screening + attention fusion" to achieve deep binding between features and network risk attributes, the accuracy of over-limit event prediction is greatly improved (model test accuracy ≥90%), enabling early identification of over-limit risks, and providing accurate basis for the classification and annotation of pre-over-limit event record text, thereby improving the accuracy of over-limit event prediction and solving the problems of traditional methods being unable to provide early warnings and the chaotic classification of pre-over-limit event record text.

[0058] S103: Input the over-limit correlation data network into the pre-trained aircraft over-limit prediction model and output the aircraft over-limit prediction results.

[0059] In this embodiment, the aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives over-limit related data from a network.

[0060] The prediction layer is used to extract the node features, edge features, adjacency features, and label association features of the over-limit associated data network; adjust the over-limit associated data network based on the node features, edge features, adjacency features, and label association features to obtain the over-limit key network; predict the aircraft over-limit prediction result based on the node features, edge features, adjacency features, label association features, and the over-limit key network; the output layer is used to output the aircraft over-limit prediction result.

[0061] The node features are extracted using basic coding algorithms (such as Label Encoding / One-Hot Encoding) or statistical aggregation methods (such as Min-Max Normalization). These node features include a node feature vector, which contains the core information corresponding to each job node. This includes the job terminal device ID (converted to a numerical code), job personnel information (personnel number code, job responsibility classification code), flight over-limit association score, and the frequency of historical over-limit events (derived from historical data statistics). The feature vector is formed by concatenating the core information corresponding to each job node to create a 1×10 node feature vector.

[0062] The encoded discrete features, flight over-limit association score, frequency of historical over-limit events, and other multi-dimensional information are concatenated into a fixed-dimensional feature vector (1×10) to ensure a uniform feature format.

[0063] Edge features are extracted using the direct feature mapping method, resulting in an edge feature vector. The specific content of these features consists of core metrics for each network edge, including edge weight, information interaction frequency, and instruction execution efficiency. These core metrics for each network edge are concatenated to form a 1×3 edge feature vector. Specifically, the values ​​for information interaction frequency and instruction execution efficiency are directly extracted from the collaborative data and combined with the calculated network edge weights to form a 1×3 edge feature vector.

[0064] Adjacency features are extracted using neighborhood statistics, represented by an adjacency feature vector. These features contain information about the distribution of network edges within each node's neighborhood, including the number of edges and the proportion of high / medium / low weight edges. For example, the neighborhood edges of each node are traversed, and the total number of edges and the proportion of high / medium / low weight edges are counted (e.g., edges with a weight ≥ 80 are high-weight edges, 50-79 are medium-weight edges, and < 50 are low-weight edges). When represented as features, the four types of neighborhood information (number of edges, proportion of high / medium / low weight edges) are concatenated into a 1×4 adjacency feature vector.

[0065] A simple probabilistic statistical method can be used to calculate the percentage of entries with a label of 1 (exceeding the limit) and a label of 0 (not exceeding the limit) in the collaborative data corresponding to each node, as well as the total frequency of historical limit-exceeding events. Then, the statistically derived percentages and frequencies are converted into values ​​in the range of 0-1, and concatenated into a 1×3 label-associated feature vector.

[0066] Adjusting the out-of-limit association data network based on node features, edge features, adjacency features, and label association features yields the out-of-limit key network, including:

[0067] (1) Based on point features, edge features, adjacency features, and label association features, the risk assessment score of the node is obtained. Specifically:

[0068] The weighting coefficients for each feature were determined as follows: After calibration using 5-fold cross-validation, the weighting coefficients for each feature were: 0.3 for node features, 0.25 for edge features, 0.2 for adjacency features, and 0.25 for label association features.

[0069] Feature normalization: The node feature vector, edge feature vector (the average edge feature of all associated edges of the node), adjacency feature vector, and label association feature vector of each node are normalized to the range of 0-1 using the min-max normalization method.

[0070] Calculate the risk assessment score: Risk assessment score = (mean of normalized node feature vector × 0.3) + (mean of normalized edge feature vector × 0.25) + (mean of normalized adjacent feature vector × 0.2) + (mean of normalized label association feature vector × 0.25) × 100. The result is rounded to one decimal place. The score ranges from 0 to 100. The higher the score, the higher the risk of the node exceeding the limit.

[0071] Node importance determination: A risk assessment score ≥80 is considered a high-risk node, 50-79 is considered a medium-risk node, and <50 is considered a low-risk node. The risk assessment score of a node is its importance score.

[0072] (2) Based on point features, edge features, adjacency features and label association features, the influence degree of network edges is evaluated, and the over-limit event labeling of each edge is determined. The over-limit event influence labeling includes Class I edges, Class II edges and Class III edges.

[0073] Specifically, this includes assessing the impact of each network edge based on node features, edge features, adjacency features, and label association features. The specific steps are as follows:

[0074] Calculating the edge influence score: A weighted summation method is used. The edge influence score = (mean of risk assessment scores of the nodes at both ends of the edge × 0.4) + (edge ​​weight × 0.3) + (weight of the label associated with the edge × 0.3). The result is rounded to one decimal place, and the score range is 0-100. The risk assessment score of each node has already been calculated above. The average of the risk assessment scores of the two nodes corresponding to an edge is the mean of the risk assessment scores of the nodes at both ends of the edge. Similarly,

[0075] Determine the impact label of the over-limit event: edge with an impact score ≥80 is classified as a Class I edge (high-risk edge, with the highest impact on the occurrence of the over-limit event), edge with an impact score of 50-79 is classified as a Class II edge (medium-risk edge, with a moderate impact on the occurrence of the over-limit event), and edge with an impact score <50 is classified as a Class III edge (low-risk edge, with the lowest impact on the occurrence of the over-limit event). The impact score label of the edge is the importance label of the edge.

[0076] (3) Generate the critical network for exceeding the limit based on the importance score of each node and the importance label of the edge. Specifically, the critical network for exceeding the limit is generated based on the importance score (risk assessment score) of each node and the importance label of the edge (impact label of the event exceeding the limit). The specific operation is as follows: retain high-risk and medium-risk nodes and first-class and second-class edges, and remove low-risk nodes and third-class edges; for the retained nodes and edges, retain all their attribute parameters (node ​​features, edge features, weights, etc.) to form the critical network for exceeding the limit, so as to ensure that the network can focus on the core collaborative relationship that induces the event exceeding the limit, reduce redundant information, and improve the efficiency and accuracy of subsequent prediction.

[0077] In the prediction layer, based on node features, edge features, adjacency features, label association features, and the over-limit key network, the aircraft over-limit prediction result is predicted. The specific steps are as follows:

[0078] The first step is to integrate and normalize the data to form a set of interconnected features:

[0079] Specific operation: Taking the over-limit critical network as the core carrier and each post node in the network as the core of association, the node features, edge features, adjacency features, and label association features are bound and integrated to achieve integrated aggregation;

[0080] (1) Integrated aggregation: For each retained node, its corresponding node feature vector, edge feature vector (the average edge feature of all associated edges of the node), adjacency feature vector, and label associated feature vector are concatenated to form a “node-edge-adjacency-label” linkage feature vector with a dimension of 1×20. Each node corresponds to a linkage feature vector, and the linkage feature vectors of all nodes are summarized to form a linkage feature set.

[0081] (2) Normalization: The min-max normalization method is used to uniformly normalize all the linkage feature vectors in the linkage feature set, and adjust the range of all parameters to 0-1. The calculation formula is: normalized value = (original value - minimum value of the feature parameter) / (maximum value of the feature parameter - minimum value of the feature parameter), eliminating the difference in magnitude between different feature parameters and ensuring the accuracy of subsequent feature selection.

[0082] The second step is to obtain the core prediction feature set based on the linked feature set: using the Pearson correlation coefficient method, the correlation coefficient between each feature in the linked feature set and the label of the event exceeding the limit is calculated, and features with a correlation coefficient ≥ 0.7 are selected to form the core prediction feature set; the correlation coefficient r is calculated as follows: , where x_i is the feature value, x1 is the feature mean, y_i is the label of the event exceeding the limit (0 or 1), and y1 is the label mean. A correlation coefficient ≥ 0.7 indicates that the feature is highly correlated with the occurrence of the event exceeding the limit and is included in the core prediction feature set; features with a correlation coefficient < 0.7 are considered redundant features and are removed to ensure the relevance of the core prediction feature set.

[0083] The third step is to integrate the core predicted feature set with the adjacency relationships of the over-limit critical network and extract the comprehensive features of the nodes:

[0084] Specific operation: Graph Convolutional Network (GCN) is used to deeply fuse the adjacency relationships between the core prediction feature set and the over-limit key network. The specific steps are as follows:

[0085] (1) Construct the adjacency matrix: Based on the nodes and edges of the critical network, construct the adjacency matrix A with a dimension of N×N (N is the number of nodes in the critical network). If there is a network edge between two nodes, then A[i][j]=1, otherwise A[i][j]=0;

[0086] (2) Feature Fusion and Extraction: The core prediction feature set is converted into a feature matrix X (with dimensions N×M, where M is the number of features in the core prediction feature set). Feature fusion is performed using the simplified layout convolution formula H=σ(A×X×W+b), where W is the weight matrix (determined through model training), b is the bias term, and σ is the activation function (using the ReLU function). The output H is the node comprehensive feature matrix, with each node corresponding to one row and a dimension of 1×M, which can comprehensively characterize the node's own over-limit risk and collaborative over-limit risk. The bias term b is a "learnable parameter" in the model training process, which does not need to be manually set or calculated in advance. Its learning process is synchronized with the weight matrix W, relying entirely on the "graph network + logistic regression" in the patent.

[0087] The fourth step is to obtain the final fused features of the nodes based on the attention fusion mechanism and the comprehensive features of the nodes:

[0088] Specific operation: A cross-layer graph attention fusion mechanism is adopted, combining the risk assessment score of nodes and the influence score of edges, to automatically assign higher weights to the features corresponding to high-risk nodes and high-risk edges. The specific steps are as follows:

[0089] (1) Calculate attention weights: For each node, calculate the attention weights between it and its neighboring nodes. Where α_ij is the attention coefficient between node i and node j, which is calculated by combining the comprehensive features of node i and node j and the influence score of the edge. α_ij = W_a × [h_i||h_j||e_ij] (W_a is the attention weight matrix, h_i and h_j are the comprehensive features of node i and node j respectively, e_ij is the influence score of the edge between node i and node j, and || represents feature concatenation).

[0090] (2) Feature weighted fusion: The comprehensive features of each node are weighted and summed with the corresponding attention weights to obtain the final fused features of the node. The formula is: h_i'=Σ(attention weight × h_j), where h_i' is the final fused feature of node i and h_j is the comprehensive feature of the neighboring nodes of node i. This step realizes the deep binding between the comprehensive features of the node and the network risk attributes.

[0091] The fifth step involves inputting the final fused features of the nodes into the logistic regression prediction model, which outputs the probability of a flight over-limit event occurring for each piece of collaborative data.

[0092] Specific operation: The final fused features of all nodes are concatenated to form a prediction feature matrix, which is then input into a pre-trained logistic regression prediction model. The model uses the sigmoid activation function to map the prediction results to the range of 0-1, and then multiplies them by 100 to obtain the probability of flight over-limit events corresponding to each collaborative data point. The probability value ranges from 0-100%, and is rounded to one decimal place. For example, if the prediction probability of a certain collaborative data point is 35.2%, it means that the probability of an over-limit event occurring in the scenario corresponding to that collaborative data point is 35.2%.

[0093] Step 6: Generate aircraft over-limit prediction results based on the probability of occurrence. Specifically: Risk Level Determination: Based on the probability of flight over-limit events, classify risks into high, medium, and low levels. The specific standards are: ≥70% probability of occurrence is high risk, 30%-69% is medium risk, and <30% is low risk. Risk Inducing Factors: Combining specific information on high-risk nodes and edges within the critical over-limit network, accurately identify the risk inducing factors. For example, if the high-risk nodes are the flight crew and air traffic control, and the high-risk edge is the network edge between them, then the risk inducing factor is "deviations in the collaborative command interaction between the flight crew and air traffic control, resulting in low execution efficiency." Early Warning Recommendations: For the identified risk inducing factors, generate targeted and actionable early warning recommendations. For example, for the above risk inducing factors, the early warning recommendation is "strengthen collaborative training between the flight crew and air traffic control, optimize command interaction processes, and improve command execution efficiency." Integrated output: The probability of occurrence, overall risk level, risk causes and early warning suggestions of each piece of collaborative data are integrated to form a complete aircraft over-limit prediction result, ensuring that the result is clear and can be used for classification and annotation of pre-over-limit event record text.

[0094] By mining the characteristics of job collaboration and constructing the critical network of exceeding limits, we can accurately locate the collaborative triggering factors of exceeding limits events, clarify the degree of influence of each job and collaborative behavior, solve the problems of unclear responsibility, slow tracking and easy error in traditional methods, and provide clear basis for the classification and labeling of pre-exceeding limit event record texts, which facilitates subsequent traceability and review, and solves the traditional problem of responsibility determination.

[0095] S104: Use the aircraft over-limit prediction results as the text classification label for the pre-over-limit event record during the aircraft's operation.

[0096] By using the aircraft over-limit prediction results as the classification label for the pre-over-limit event record text, the problem of the lack of specificity in the classification of existing pre-over-limit event record texts is solved. This achieves accurate classification and labeling of pre-over-limit event record texts, which facilitates subsequent retrieval, review analysis, and model iteration optimization of pre-over-limit event record texts, and realizes a deep binding between pre-over-limit event record texts and over-limit prediction.

[0097] After pre-labeling the over-limit events during aircraft operation (which can also be seen as pre-labeling of over-limit events in the historical over-limit association collaborative data list), the pre-over-limit event record text is classified and labeled and sent to the management center so that the management center can obtain the over-limit estimation information of the aircraft operation in a timely manner, thereby taking timely measures to avoid over-limit events, reducing the probability of over-limit events, and improving civil aviation flight safety.

[0098] By categorizing and labeling pre-excessive event records and pushing them to the management center in real time, timely transmission of categorization labels is achieved. This allows managers to quickly grasp the current flight exceedance risk situation and take preventive measures in advance. This solves the problems of untimely transmission of categorization labels and low management efficiency in traditional methods, thereby improving management efficiency.

[0099] As an optional embodiment: This embodiment takes the operation of flight CA1234 (route: Beijing-Shanghai, aircraft type: Boeing 737, flight duration: 2 hours 30 minutes) as an example. The specific implementation process of the deep learning-based flight over-limit event record text classification method is as follows:

[0100] Step 1: Obtain a list of historical out-of-limit collaborative data.

[0101] Through the Civil Aviation Flight Data Management System, historical out-of-limit correlation and collaboration data were collected from relevant positions (flight crew, air traffic control, dispatch, and ground staff) on similar routes (Beijing-Shanghai) and similar aircraft types (Boeing 737) within the three months prior to the current flight of flight CA1234. A total of 1200 valid data entries were collected. After deduplication, noise reduction, and completion preprocessing (missing execution instruction efficiency was completed using the average of data of the same type and position, and abnormal data was removed using the 3σ criterion), a standardized historical out-of-limit correlation and collaboration data list was formed. The core data is as follows:

[0102] 1. Terminal device ID information for each position: Flight crew (FLT-001), Air traffic control (ATC-001), Dispatch (DIS-001), Ground crew (GRD-001);

[0103] 2. Job titles and personnel information: Flight crew (personnel number FL-001, job responsibilities: executing flight instructions and adjusting flight attitude), air traffic control (personnel number AT-001, job responsibilities: issuing altitude and speed instructions), dispatch (personnel number DI-001, job responsibilities: planning flight routes), ground crew (personnel number GR-001, job responsibilities: ground support coordination);

[0104] 3. Collaborative Data:

[0105] (1) Information interaction records: 320 interactions between the flight crew and air traffic control (all altitude and speed commands, with an average response time of 25 seconds); 180 interactions between the flight crew and dispatch (all route adjustment commands, with an average response time of 35 seconds); 210 interactions between air traffic control and dispatch (all command connection interactions, with an average response time of 30 seconds); no effective interactions between ground staff and other positions.

[0106] (2) Record of executed instructions: Flight crew instruction qualification rate 92%, average execution time 28 seconds; Air traffic control instruction qualification rate 95%, average execution time 22 seconds; Dispatch instruction qualification rate 88%, average execution time 38 seconds;

[0107] (3) Information exchange frequency and execution efficiency (out of 100): Flight crew - air traffic control (6 times / hour, 85 points); Flight crew - dispatch (3 times / hour, 90 points); Air traffic control - dispatch (4 times / hour, 88 points); Ground crew (0 times / hour, 80 points).

[0108] 4. Flight over-limit correlation rating (average): Flight crew 8.25 points, air traffic control 8.25 points, dispatch 4.5 points, ground crew 2.5 points.

[0109] 5. Over-limit event tags: Flight crew-air traffic control coordination data (tag 1: 96 records, tag 0: 224 records); Flight crew-dispatch coordination data (tag 1: 18 records, tag 0: 162 records); Air traffic control-dispatch coordination data (tag 1: 31 records, tag 0: 179 records); Ground crew no coordination data (tag set to 0).

[0110] Step 2: Constructing an over-limit correlated data network:

[0111] Based on the above list of historical out-of-limit association and collaboration data, the out-of-limit association data network is constructed according to the following steps, with the core calculation process as follows:

[0112] 1. Node construction: The network nodes are based on four positions: flight crew, air traffic control, dispatch, and ground crew. Each node is associated with the corresponding terminal ID, personnel information, average value of flight over-limit association labels, and historical over-limit frequency (12 times for flight crew, 10 times for air traffic control, 5 times for dispatch, and 0 times for ground crew).

[0113] 2. Network edge construction: Based on information interaction records, three network edges are constructed between the flight crew and air traffic control, the flight crew and dispatch, and air traffic control and dispatch. No edge is constructed for ground staff who have no effective interaction.

[0114] 3. Network edge weight calculation (using min-max normalization + weighted summation, weighting coefficients: information interaction frequency 0.15, execution command efficiency 0.25, flight over-limit association weight 0.25, tag association weight 0.35):

[0115] (1) Formula for calculating the correlation weight of flight over-limit:

[0116] Flight crew - air traffic control side: (8.25 + 8.25) / 2 × 10 = 82.5 points;

[0117] Flight crew - dispatch edge: (8.25 + 4.5) / 2 × 10 = 63.75 points;

[0118] Air traffic control - dispatch edge: (8.25+4.5) / 2 ×10 = 63.75 points.

[0119] (2) Formula for calculating tag association weight:

[0120] Flight crew - air traffic control side: (96 / 320)×100 = 30.0 points;

[0121] Flight crew - dispatch edge: (18 / 180) × 100 = 10.0 points;

[0122] Air traffic control - dispatch edge: (31 / 210)×100 ≈14.76 points.

[0123] (3) Normalization formula:

[0124] Parameter extreme values: information interaction frequency (1-10 times / hour), execution command efficiency (70-95 points), flight over-limit association weight (25-85 points), tag association weight (0-40 points), then:

[0125] Flight crew-air traffic control side: Information exchange frequency (6-1) / (10-1)×100≈55.6 points; Execution command efficiency (85-70) / (95-70)×100=60.0 points; Flight over-limit association weight (82.5-25) / (85-25)×100≈95.8 points; Tag association weight (30.0-0) / (40-0)×100=75.0 points.

[0126] Flight crew-dispatch side: Information interaction frequency (3-1) / (10-1)×100≈22.2 points; Execution command efficiency (90-70) / (95-70)×100=80.0 points; Flight over-limit association weight (63.75-25) / (85-25)×100≈64.58 points; Tag association weight (10.0-0) / (40-0)×100=25.0 points.

[0127] Air traffic control-dispatch side: Information interaction frequency (4-1) / (10-1)×100≈33.3 points; Execution instruction efficiency (88-70) / (95-70)×100=72.0 points; Flight over-limit association weight (63.75-25) / (85-25)×100≈64.58 points; Tag association weight (14.76-0) / (40-0)×100≈36.90 points.

[0128] Final weight calculation formula:

[0129]

[0130] Wherein, A is the normalized information interaction frequency (the value of the information interaction frequency of the corresponding network edge after min-max normalization, such as the value of A for the flight crew-air traffic control edge being 55.6 points); B is the normalized execution command efficiency (the value of the execution command efficiency of the corresponding network edge after min-max normalization, such as the value of B for the flight crew-air traffic control edge being 60.0 points); C is the normalized flight over-limit association weight (the value of the flight over-limit association weight of the corresponding network edge after min-max normalization, such as the value of C for the flight crew-air traffic control edge being 95.8 points); and D is the normalized label association weight (the value of the label association weight of the corresponding network edge after min-max normalization, such as the value of D for the flight crew-air traffic control edge being 75.0 points).

[0131] The final constructed super-relational data network consists of 4 nodes and 3 edges, with edge weights of 73.54, 48.23, and 52.06 respectively.

[0132] Step 3: Input the aircraft over-limit prediction model and output the over-limit prediction results.

[0133] The patented hybrid model of "graph network + logistic regression" (pre-trained: 10,000 training data points, 7:2:1 training / validation / test set partitioning, Adam optimizer, learning rate 0.001, 100 iterations, validation set accuracy 92%, loss value 0.08) was input into the aforementioned over-limit correlation data network. The specific execution process is as follows:

[0134] 1. Input layer: Parses network data, converts node information, edge information, edge weights, etc. into tensor format that the model can recognize, and feeds it into the recognition layer;

[0135] 2. Recognition Layer:

[0136] (1) Feature extraction:

[0137] Node features: The terminal ID, personnel number, and job responsibility code of each position are combined with the flight over-limit association annotation and historical over-limit frequency to form a 1×10 feature vector (Example: Flight crew node feature vector: [1,1,1,8.25,12,0,0,0,0,0], encoding rules: FLT-001 is encoded as 1, FL-001 is encoded as 1, and flight crew job responsibility code is 1).

[0138] Edge features: Extract the edge weight, information interaction frequency, and execution command efficiency of each edge, and concatenate them into a 1×3 feature vector (Example: Flight crew-air traffic control edge feature vector: [73.54,6,85]).

[0139] Adjacency features: Count the number of neighborhood edges of each node and the proportion of high / medium / low weight edges (weight ≥80 is high, 50-79 is medium, <50 is low), and concatenate them into a 1×4 feature vector (Example: Flight crew node adjacency feature vector: [2,0,1,1], that is, 2 neighborhood edges, 0 high weight edges, 1 medium weight edge, and 1 low weight edge).

[0140] Tag association features: Statistically calculate the proportion of collaborative data labeled 1 and 0 for each node and the historical frequency of exceeding limits, and concatenate them into a 1×3 feature vector (Example: Flight crew node tag association feature vector: [0.25, 0.75, 12]).

[0141] (2) Network Adjustment: Based on the above characteristics, calculate the node risk assessment score and edge impact score to generate an over-limit critical network:

[0142] Formula for calculating node risk assessment score:

[0143] Risk assessment score

[0144] E / F / G / H are the normalized means of node features, variable features, adjacency features, and label association features, respectively. Example:

[0145] Flight crew: (0.82×0.3)+(0.75×0.25)+(0.60×0.2)+(0.70×0.25)×100≈76.6 points;

[0146] Empty pipe: (0.80×0.3)+(0.72×0.25)+(0.58×0.2)+(0.68×0.25)×100≈74.1 points;

[0147] Dispatch: (0.45×0.3)+(0.50×0.25)+(0.42×0.2)+(0.40×0.25)×100≈45.4 points;

[0148] Ground staff: (0.25×0.3)+(0.0×0.25)+(0.0×0.2)+(0.0×0.25)×100≈7.5 points.

[0149] Formula for calculating the degree of influence of an edge:

[0150] (J) 3) + (K) 3)

[0151] I1 represents the risk assessment score of a node at one end of the network edge; I2 represents the risk assessment score of a node at the other end of the network edge; J represents the final weight of the network edge; and K represents the label association weight of the network edge. Example:

[0152] Flight crew - air traffic control side: ((76.6+74.1) / 2×0.4)+(73.54×0.3)+(30.0×0.3)≈60.2 points;

[0153] Flight crew - dispatch edge: ((76.6+45.4) / 2×0.4)+(48.23×0.3)+(10.0×0.3)≈41.9 points;

[0154] Air traffic control - dispatch edge: ((74.1+45.4) / 2×0.4)+(52.06×0.3)+(14.76×0.3)≈44.8 points.

[0155] Over-limit critical network: retain medium- and high-risk nodes (flight crew, air traffic control) and medium- and high-risk edges (flight crew-air traffic control edge), and remove dispatch, ground crew and the other two low-risk edges.

[0156] Prediction operation:

[0157] Aggregation: The node features and edge features within the critical network are concatenated into a 1×20 linked feature vector, which is then normalized by min-max. Then, core feature selection is performed: the Pearson correlation coefficient algorithm is used to select features with r≥0.7 to form the core prediction feature set.

[0158] Feature fusion: Using the graph convolution formula: H=σ(A×X×W+b), where σ is the ReLU function, and W and b are parameters learned during model training, core features and network adjacency relationships are fused to extract comprehensive node features;

[0159] Attention Fusion: Formula for Calculating Attention Weights: , Let be the attention coefficient between node i and node j. The final fused feature of the nodes is obtained by weighted fusion based on attention weights.

[0160] Probability Prediction: Input the logistic regression model, and output the probability of exceeding limits in the coordinated data through the sigmoid activation function. Finally, the prediction result of exceeding limits for flight CA1234 is obtained: the probability of exceeding limits is 62.3%, the risk level is medium risk, the risk cause is "the deviation in the interaction of coordinated instructions between the flight crew and air traffic control, with high edge weights", and the warning suggestion is "strengthen the control of the interaction of coordinated instructions between the two".

[0161] 3. Output layer: Output the standardized over-limit prediction results to ensure that managers can clearly obtain core risk information.

[0162] Step 4: Use the prediction results as the text classification and labeling for pre-exceedance event records.

[0163] The pre-exclusion event log text (including suspected altitude exceedance data, abnormal crew coordination records, etc.) generated in real time during the flight of flight CA1234 was collected. The exceedance prediction results output from the third step were used as classification and labeling information, with the labeling format as "[Pre-exclusion classification labeling] Risk level: medium risk; Probability of exceedance occurrence: 62.3%; Risk cause: Deviation in the interaction of flight crew and air traffic control coordination instructions, with high edge weight; Warning suggestion: Strengthen the control of the interaction of coordination instructions between the two." At the same time, the pre-exclusion event log text was classified and archived according to the medium risk level, thus completing the execution of the entire method.

[0164] In summary, the entire process of "data acquisition - network construction - feature extraction - prediction and classification" has been fully realized, accurately completing the classification and labeling of pre-excessive event record texts, and simultaneously achieving accurate prediction of flight excess risk.

[0165] As another optional implementation, this application provides a deep learning-based method for classifying flight over-limit event log texts, including:

[0166] Obtain a historical list of over-limit association and collaboration data for all relevant positions involved in the current flight's operation. This list includes the position's terminal device ID, position name, personnel information, collaboration data, flight over-limit association annotations, and over-limit event tags. The collaboration data includes records of information interaction between the position's terminal device and other positions' terminal devices, records of executed instructions, the frequency of information interaction between the position's terminal device and other terminal devices, and the efficiency of executed instructions. Based on this historical list of over-limit association and collaboration data, construct an over-limit association data network. The over-limit association data network is input into a pre-trained aircraft over-limit prediction model, which outputs the aircraft over-limit prediction result. The aircraft over-limit prediction model includes an input layer, an identification layer, and an output layer. The input layer receives the over-limit association data network. The identification layer extracts the node features, edge features, adjacency features, and label association features of the over-limit association data network. Based on the node features, edge features, adjacency features, and label association features, the over-limit association data network is adjusted to obtain the over-limit key network. Based on the node features, edge features, adjacency features, label association features, and the over-limit key network, the aircraft over-limit prediction result is predicted. The output layer outputs the aircraft over-limit prediction result. The aircraft over-limit prediction result is used as the text classification label for the pre-over-limit event record of the aircraft's operation process.

[0167] Optionally, the method also includes classifying and labeling the pre-excess event record text and sending it to the management center, so that managers can keep abreast of the current flight excess risk situation and carry out control work in a timely manner.

[0168] Optionally, based on the historical list of out-of-limit correlation and collaboration data, an out-of-limit correlation data network is constructed, including:

[0169] The network of data related to flight over-limit association is based on each position during the current flight's operation. Each node records the corresponding position name and is associated with the terminal device ID, personnel information, and flight over-limit association label for that position.

[0170] Network edges are constructed based on information interaction records in collaborative data. If the terminal devices of two positions have information interaction records (including interaction records related to flight over-limit), a network edge is constructed between the nodes corresponding to these two positions; if the terminal devices of two positions have no information interaction records, no network edge is constructed.

[0171] Based on collaborative data, flight over-limit association annotations, and over-limit event occurrence labels, the weight of each network edge is obtained through quantitative calculation. The weight value is used to characterize the degree of anomaly of the collaborative edge and the strength of its association with the over-limit event.

[0172] Optionally, based on collaborative data, flight over-limit association annotations, and over-limit event occurrence labels, the weights of network edges are obtained, including:

[0173] The first step is to calculate the flight over-limit association weight and the label association weight respectively. The flight over-limit association weight is calculated based on the flight over-limit association annotation data. Specifically, the average value of the flight over-limit association annotations for the corresponding two nodes is taken and then multiplied by 10 to obtain a quantitative score, which is used to characterize the inherent association strength between the collaborative edge and the over-limit event. The label association weight is calculated based on the over-limit event occurrence label data. The historical over-limit label ratio corresponding to the collaborative edge is statistically analyzed and then mapped to a quantitative score of 0-100, which is used to characterize the probability that the collaborative edge has historically induced an over-limit event.

[0174] The second step is to normalize the information interaction frequency and execution command efficiency in the collaborative data, as well as the flight over-limit association weight and tag association weight calculated in the first step, and adjust the value range of all parameters to 0-100 points to avoid the impact of parameter magnitude differences on the accuracy of weight calculation.

[0175] The third step involves weighted summation of the normalized information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight to obtain the final weights of the network edges. The weighting coefficients are determined through 5-fold cross-validation calibration, with execution command efficiency and flight over-limit association weight each accounting for 25%, information interaction frequency and label association weight each accounting for 15%, and the remaining 20% ​​being an abnormal interaction correction term (the correction term is 0 when there is no abnormal interaction).

[0176] Optionally, the excess-limit association data network can be adjusted based on node features, edge features, adjacency features, and label association features to obtain the excess-limit key network, including:

[0177] The first step is node risk assessment: Based on the node features, edge features, adjacency features, and label association features extracted from the identification layer, a weighted summation method is used to calculate the risk assessment score (i.e., node importance score) for each node. The weight allocation is consistent with the weighting coefficient for calculating the network edge weights. The higher the score, the higher the risk of the node exceeding the limit. The specific calculation formula is: Node risk assessment score = (average value of flight over-limit associated labels corresponding to the node × 25%) + (average weight of the node's associated edges × 25%) + (proportion of high-risk edges in the node's neighborhood × 15%) + (proportion of historical over-limit labels corresponding to the node × 15%).

[0178] The second step is to assess the impact of network edges: Based on node features, edge features, adjacency features, and label association features, the impact of each network edge is assessed. By calculating the label score of the edge, the network edges are divided into three categories: Category I edges (high-risk edges, label score ≥ 80 points), Category II edges (medium-risk edges, label score 40-79 points), and Category III edges (low-risk edges, label score < 40 points). Among them, Category I edges have the highest impact on the occurrence of over-limit events, and Category III edges have the lowest impact.

[0179] The third step is to generate the critical network for exceeding the limit: Based on the risk assessment score (node ​​importance score) of each node and the impact label of the event exceeding the limit for each edge (edge ​​importance label), high-risk and medium-risk nodes (risk assessment score ≥ 40 points) and Class I and Class II edges are retained, while low-risk nodes (risk assessment score < 40 points) and Class III edges are removed to generate the critical network for exceeding the limit, ensuring that the network focuses on the core collaborative relationships that induce the event exceeding the limit.

[0180] Optionally, based on node features, edge features, adjacency features, label association features, and the over-limit key network, the aircraft over-limit prediction result is predicted, including:

[0181] The first step is feature aggregation and normalization: The node features, edge features, adjacency features, and label association features of the critical network exceeding limits are aggregated in an integrated manner. Taking a node as the core, the corresponding edge features (weights of associated edges, average interaction indicators), adjacency features (distribution ratio of high / medium risk edges within the neighborhood), and label association features (historical statistics of exceeding limits labels) are bound together to form a set of linked features. The min-max normalization method is used to normalize all parameters in the linked feature set to the range of 0-1. The calculation formula is: Normalized value = (Original value - Minimum value of the parameter) / (Maximum value of the parameter - Minimum value of the parameter), eliminating differences in parameter magnitude.

[0182] The second step is core feature selection: Based on Pearson correlation coefficient analysis, features with a correlation of ≥0.7 with the label of the occurrence of the over-limit event are selected to form the core prediction feature set. Redundant features with low correlation are removed to improve prediction efficiency.

[0183] The third step is node comprehensive feature extraction: a simplified layout convolution operation is adopted to deeply integrate the adjacency relationship between the core prediction feature set and the over-limit key network. Without the need for hierarchical iterative calculation, the collaborative risk association between nodes and neighboring nodes is directly mined, and the node comprehensive features that can comprehensively characterize the node's own over-limit risk and collaborative over-limit risk are output.

[0184] The fourth step is cross-layer attention feature fusion: through the cross-layer graph attention fusion mechanism, the risk assessment scores of nodes and the labeled scores of edges in the over-limit critical network are combined to assign higher weights to the features corresponding to high-risk nodes and high-risk edges, so as to achieve deep binding between the comprehensive features of nodes and the network risk attributes, and output the final fused features of nodes.

[0185] Step 5, Over-limit probability prediction: Input the final fused features of the nodes into the preset logistic regression prediction model, combine the real-time risk attributes of nodes and edges in the over-limit key network, and output the probability of flight over-limit events corresponding to each piece of collaborative data (value range 0-100%).

[0186] Step 6: Integrate Prediction Results: Based on the probability of flight over-limit events, classify risk levels into high, medium, and low (high risk: probability ≥70%, medium risk: 30%-69%, low risk: <30%). Combine the specific information of high-risk nodes and edges within the critical over-limit network to accurately identify the risk triggers (such as deviations in the execution of instructions for a certain position, abnormal interaction frequency between two positions, etc.). Generate targeted and actionable early warning suggestions (such as strengthening the review of instructions for that position, optimizing the interaction process between positions, etc.). Integrate the probability of occurrence, risk level, risk triggers, and early warning suggestions to form a complete aircraft over-limit prediction result.

[0187] This application also provides a deep learning-based flight over-limit event record text classification system for performing the aforementioned deep learning-based flight over-limit event record text classification method. The system includes:

[0188] The acquisition module is used to acquire a historical list of over-limit associated collaborative data for all relevant positions involved in the current flight's operation. This historical over-limit associated collaborative data list includes position terminal device ID information, position name, personnel information, collaborative data, flight over-limit association annotations, and over-limit event occurrence tags. The collaborative data includes information interaction records between the position's terminal device and other position terminal devices, execution command records, the frequency of information interaction between the position's terminal device and other terminal devices, and execution command efficiency. The network construction module is used to construct an over-limit associated data network based on the historical over-limit associated collaborative data list.

[0189] The prediction module is used to input the over-limit correlation data network into a pre-trained aircraft over-limit prediction model and output the aircraft over-limit prediction result. The aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives the over-limit correlation data network; the prediction layer extracts the node features, edge features, adjacency features, and label association features of the over-limit correlation data network; adjusts the over-limit correlation data network based on the node features, edge features, adjacency features, and label association features to obtain the over-limit key network; predicts the aircraft over-limit prediction result based on the node features, edge features, adjacency features, label association features, and the over-limit key network; and the output layer outputs the aircraft over-limit prediction result.

[0190] The classification and labeling module is used to classify and label the pre-excess event record text of the aircraft operation process based on the aircraft excess prediction results.

[0191] By adopting the above technical solutions, a closed-loop process of "historical data acquisition → network construction → model prediction → pre-excess text classification and labeling → push management" is formed, realizing a closed loop of excess risk and pre-excess text management. This not only enables the early prediction of excess events, but also allows for the efficient reuse of pre-excess event record texts through classification and labeling, providing strong support for subsequent excess risk prevention and management optimization, and improving the level of civil aviation flight safety and the efficiency of pre-excess event record text management.

Claims

1. A deep learning-based text classification method for flight over-limit event records, characterized in that, include: Obtain a list of historical over-limit association and collaboration data for all relevant positions involved in the current flight's operation; the list of historical over-limit association and collaboration data includes position terminal device ID information, position name, position personnel information, collaboration data, flight over-limit association annotations, and over-limit event occurrence tags; the collaboration data includes information interaction records between the position's terminal device and other position terminal devices, records of executed instructions, information interaction frequency between the position's terminal device and other terminal devices, and execution instruction efficiency; Construct an over-limit association data network based on a historical list of over-limit associated collaborative data; Input the over-limit correlation data network into the pre-trained aircraft over-limit prediction model and output the aircraft over-limit prediction results; The aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives the over-limit associated data network; the prediction layer extracts the node features, edge features, adjacency features, and label association features of the over-limit associated data network; the over-limit associated data network is adjusted based on the node features, edge features, adjacency features, and label association features to obtain the over-limit key network; and the aircraft over-limit prediction result is predicted based on the node features, edge features, adjacency features, label association features, and the over-limit key network. The output layer is used to output the aircraft over-limit prediction results; The aircraft over-limit prediction results are used as the classification and labeling text of the pre-over-limit event records during the aircraft's operation. Among them, the excess-limit association data network is adjusted based on node features, edge features, adjacency features, and label association features to obtain the excess-limit key network, including: The risk assessment score of a node is obtained by adjusting the node features, edge features, adjacency features, and label association features. Based on point features, edge features, adjacency features, and label association features, the degree of influence of network edges is evaluated, and the out-of-limit event labeling for each edge is determined. The out-of-limit event influence labeling includes Class I edges, Class II edges, and Class III edges. An over-critical network is generated based on the importance score of each node and the importance label of each edge.

2. The deep learning-based text classification method for flight over-limit event records according to claim 1, characterized in that, The method also includes classifying and labeling the pre-exceedance event record text and sending it to the management center.

3. The deep learning-based text classification method for flight over-limit event records according to claim 1, characterized in that, Based on the list of out-of-limit collaborative data, construct an out-of-limit data network, including: Each position during the current flight's operation is taken as a node in the over-limit association data network. Each node is recorded as the corresponding position name. Network edges are constructed based on the information interaction records in the collaborative data. The weights of the network edges are obtained based on the collaborative data, flight over-limit association labels, and over-limit event occurrence labels.

4. The deep learning-based text classification method for flight over-limit event records according to claim 3, characterized in that, Based on collaborative data, flight over-limit association annotations, and over-limit event occurrence labels, the weights of network edges are obtained, including: The flight over-limit association weight and the label association weight are obtained based on the flight over-limit association annotation and the over-limit event occurrence label, respectively; The information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are normalized respectively. The normalized information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are then weighted and summed to obtain the weight of the network edge.

5. The deep learning-based text classification method for flight over-limit event records according to claim 1, characterized in that, Based on node features, edge features, adjacency features, label association features, and the over-limit key network, the aircraft over-limit prediction results are generated, including: Node features, edge features, adjacency features, and label association features are aggregated and normalized to form a set of linked features; The core prediction feature set is obtained based on the linked feature set; The core predictive feature set is integrated with the adjacency relationship of the over-limit critical network, and the comprehensive features of the nodes are extracted. Based on the attention fusion mechanism, the final fusion features of nodes are obtained based on the comprehensive features of nodes; The final fused features of the nodes are input into the logistic regression prediction model, which outputs the probability of flight over-limit events corresponding to each piece of collaborative data. The prediction results for aircraft exceeding limits are generated based on the probability of occurrence.

6. A deep learning-based text classification system for flight over-limit event records, characterized in that, include: The acquisition module is used to acquire a list of historical over-limit association and collaboration data for all relevant positions involved in the current flight's operation. The list of historical over-limit association and collaboration data includes the position's terminal device ID information, position name, personnel information, collaboration data, flight over-limit association annotations, and over-limit event occurrence tags. The collaboration data includes information interaction records between the position's terminal device and other position's terminal devices, records of executed instructions, the frequency of information interaction between the position's terminal device and other terminal devices, and the efficiency of executed instructions. The network construction module is used to build an over-limit association data network based on a historical list of over-limit association collaborative data; The prediction module is used to input the over-limit correlation data network into a pre-trained aircraft over-limit prediction model and output the aircraft over-limit prediction result. The aircraft over-limit prediction model includes an input layer, a prediction layer, and an output layer. The input layer receives the over-limit correlation data network; the prediction layer extracts node features, edge features, adjacency features, and label association features from the over-limit correlation data network; adjusts the over-limit correlation data network based on these features to obtain the over-limit key network; and predicts the aircraft over-limit based on the node features, edge features, adjacency features, label association features, and the over-limit key network. Results; The output layer is used to output the aircraft over-limit prediction results; Among them, the over-limit associated data network is adjusted based on node features, edge features, adjacency features, and label association features to obtain the over-limit key network, including: adjusting based on point features, edge features, adjacency features, and label association features to obtain the risk assessment score of the node; evaluating the degree of influence of the network edges based on point features, edge features, adjacency features, and label association features to determine the over-limit event label of each edge, including Class I edges, Class II edges, and Class III edges; generating the over-limit key network based on the importance score of each node and the importance label of the edge; The classification and labeling module is used to classify and label the pre-excess event record text of the aircraft operation process based on the aircraft excess prediction results.

7. The deep learning-based flight over-limit event record text classification system according to claim 6, characterized in that, The system also includes: The sending module is used to classify and label the pre-exceedance event record text and send it to the management center.

8. The deep learning-based flight over-limit event record text classification system according to claim 6, characterized in that, Based on the list of out-of-limit collaborative data, construct an out-of-limit data network, including: Each position during the current flight's operation is taken as a node in the over-limit association data network. Each node is recorded as the corresponding position name. Network edges are constructed based on the information interaction records in the collaborative data. The weights of the network edges are obtained based on the collaborative data, flight over-limit association labels, and over-limit event occurrence labels.

9. The deep learning-based flight over-limit event record text classification system according to claim 6, characterized in that, Based on collaborative data, flight over-limit association annotations, and over-limit event occurrence labels, the weights of network edges are obtained, including: The flight over-limit association weight and the label association weight are obtained based on the flight over-limit association annotation and the over-limit event occurrence label, respectively; The information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are normalized respectively. The normalized information interaction frequency, execution command efficiency, flight over-limit association weight, and label association weight are then weighted and summed to obtain the weight of the network edge.

Images