Heavy landing risk prediction method, heavy landing risk model training method, heavy landing risk prediction system and electronic equipment

Abstract

The invention discloses a training method of a heavy landing risk prediction model, a heavy landing risk prediction method and system and electronic equipment, and relates to the technical field of flight safety technologies, and the method comprises the steps: screening out heavy landing feature data based on historical flight landing data of a pilot; performing data expansion on the heavy landing feature data to generate expanded heavy landing feature data; classifying the expanded heavy landing feature data according to a time scale, and generating a training data set of an initial heavy landing risk prediction model corresponding to a pilot; wherein the training data set comprises expanded heavy landing feature data corresponding to each time scale and landing labels corresponding to the heavy landing feature data; and training an initial heavy landing risk prediction model based on the training data set to generate a heavy landing risk prediction model corresponding to the pilot. The method can achieve the effects of improving the risk prediction capability, and improving the pertinence, adaptability and practicability of risk prediction.

Description

Technical Field

[0001] This invention relates to the field of flight safety technology, and in particular to a method for predicting hard landing risks, a training method for a hard landing risk model, a hard landing risk prediction system, and electronic equipment. Background Technology

[0002] According to civil aviation safety operating standards, aircraft landings are classified into two categories: normal landings and hard landings. A hard landing occurs when, during the touchdown phase, the vertical load on the aircraft is excessive or the sinking rate exceeds a set threshold. Once this happens, minor incidents may cause wear or slight deformation of components such as the landing gear and tires, leading to flight delays; more serious incidents may result in damage to the fuselage's load-bearing structure, landing gear deformation, or even breakage; in extreme cases, it can even cause secondary accidents such as the aircraft veering off the runway, seriously threatening civil aviation flight safety. Therefore, it is necessary to conduct precise analysis of the flight status during the aircraft's approach phase to provide the crew with real-time risk warnings, enabling them to adjust their operations in a timely manner and avoid hard landing events from the outset.

[0003] Currently, the prediction of hard landing risk is either based on post-event analysis or by using a traditional machine learning prediction model. This method involves labeling historical landing data into two categories: normal and hard landing, extracting approach phase feature parameters to train a classification model, and then predicting the probability of hard landing risk for a single approach.

[0004] However, all of the aforementioned existing technologies suffer from weak risk warning capabilities. Summary of the Invention

[0005] This invention aims to address one of the technical problems in related technologies to a certain extent. To this end, this invention provides a method for predicting hard landing risks, a method for training a hard landing risk model, a hard landing risk prediction system, and an electronic device for executing the prediction and training methods. This invention can improve risk prediction capabilities, enhance the relevance, adaptability, and practicality of risk prediction.

[0006] To achieve the above objectives, as a first aspect of the present invention, a training method for a hard landing risk prediction model is provided, wherein the training method includes: Based on the pilot's historical flight and landing data, select re-landing characteristic data; The hard landing feature data is augmented to generate augmented hard landing feature data; The expanded hard landing feature data is classified according to time scale to generate a training dataset for the initial hard landing risk prediction model corresponding to the pilot. The training dataset includes the augmented relanding feature data corresponding to each time scale and the landing labels corresponding to the relanding feature data; The initial hard landing risk prediction model is trained based on the training dataset to generate a hard landing risk prediction model corresponding to the pilot.

[0007] In some feasible implementations, the filtering of relanding characteristic data based on the pilot's historical flight and landing data includes: By filtering the various flight and landing parameters of the historical flight and landing data, re-landing spatial feature data is generated. By filtering the various flight and landing parameters of the historical flight and landing data, re-landing time feature data is generated. The historical flight and landing data includes multiple flight and landing parameters; Each flight landing parameter of the hard landing feature data is greater than or equal to a preset physical parameter threshold; the hard landing feature data includes hard landing spatial feature data and hard landing temporal feature data.

[0008] In some feasible implementations, the step of filtering the various flight landing parameters of the historical flight landing data to generate relanding spatial feature data includes: The contribution of each flight landing parameter to the smaller prediction error of the classification model is determined based on the classification model. The flight and landing parameters are sorted according to their contribution, and the first to Nth sorted flight and landing parameters are selected. Based on the sorted first to Nth flight landing parameters, re-landing spatial feature data is generated, wherein N is greater than 1 and the contribution of the Nth flight landing parameter is greater than or equal to a preset contribution.

[0009] In some feasible implementations, the step of filtering the various flight landing parameters of the historical flight landing data to generate relanding time feature data includes: The target entropy of each flight and landing parameter within the set flight and landing time period is determined based on the number of times each flight and landing parameter appears within the set flight and landing time period. Based on the target entropy, sort the flight and landing parameters, and select the first to Mth sorted flight and landing parameters; Based on the first to M flight landing parameters, re-landing time characteristic data is generated; Where M is greater than 1 and the target entropy of the Mth flight landing parameter is greater than or equal to the preset entropy.

[0010] In some feasible implementations, the step of augmenting the hard landing feature data to generate augmented hard landing feature data includes: Based on feature engineering, multiple statistical dimensions of the hard landing spatial feature data are determined, and derived hard landing spatial feature data corresponding to each statistical dimension are generated. Based on the hard landing spatial feature data and the corresponding derived hard landing spatial feature data, expanded hard landing spatial feature data is generated; Based on feature engineering, multiple statistical dimensions of the hard landing time feature data are determined, and derived hard landing time feature data corresponding to each statistical dimension are generated. Based on the hard landing time feature data and the corresponding derived hard landing time feature data, extended hard landing time feature data is generated; The expanded hard landing feature data includes expanded hard landing spatial feature data and expanded hard landing temporal feature data.

[0011] In some feasible implementations, the initial relanding risk prediction model includes: a cascaded feature extraction module and a feature fusion classification module. The feature extraction module includes multiple parallel feature extraction sub-modules corresponding to the time scale; The feature extraction submodule is used to extract the spatiotemporal features of the augmented re-landing feature data corresponding to the time scale; The feature fusion classification module is used to fuse the spatiotemporal features corresponding to each extracted time scale and output the classification result. The classification results include hard landing and normal landing.

[0012] In some feasible implementations, the extraction of spatiotemporal features of the augmented relanding feature data corresponding to the time scale includes: One-hot encoding is performed on the augmented hard landing time feature data and the augmented hard landing spatial feature data respectively to generate augmented hard landing time feature vector and augmented hard landing spatial feature vector; The dimensions of the extended hard landing time feature vector and the extended hard landing space feature vector are reconstructed respectively to generate a fixed-dimensional extended hard landing time feature structure and an extended hard landing space feature structure. A time-series value encoder based on a spatiotemporal encoder extracts features from the augmented relanding time feature structure to generate time-series features; The measurement encoder based on the spatiotemporal encoder extracts features from the extended hard landing spatial feature structure to generate spatial features; The temporal features and spatial features are merged to generate merged features; The merged features are weighted and aggregated using an attention mechanism to generate spatiotemporal features. The spatiotemporal features include: contextual semantic features corresponding to the time scale.

[0013] As a second aspect of the present invention, a method for predicting the risk of a hard landing is provided, wherein the method for predicting the risk of a hard landing includes: Obtain pilot information and flight / landing data for the current flight; The flight and landing data are filtered to generate hard landing feature data; Expand the hard landing feature data to generate expanded hard landing feature data; The augmented hard landing feature data is input into a model trained by the training method of the hard landing risk prediction model provided in the first aspect of the present invention.

[0014] As a third aspect of the present invention, a hard landing risk prediction system is also provided, comprising: The filtering module is used to filter out hard landing characteristic data from the pilot's historical flight landing data; An expansion module is used to expand the hard landing feature data to generate expanded hard landing feature data; The generation module is used to classify the expanded hard landing feature data according to time scales and generate a training dataset for the initial hard landing risk prediction model corresponding to the pilot; wherein, the training dataset includes expanded hard landing feature data corresponding to each time scale and landing labels corresponding to the hard landing feature data; The training module is used to train the initial hard landing risk prediction model using the training dataset to generate a hard landing risk prediction model corresponding to the pilot. The prediction module is used to input the expanded hard landing feature data into the hard landing risk prediction model corresponding to the acquired pilot information, and generate the hard landing risk prediction result for the current flight, wherein the hard landing risk prediction model is a model generated by the training module.

[0015] Furthermore, as a fourth aspect of the present invention, an electronic device is provided, comprising: One or more processors; A memory having stored one or more computer programs that, when executed by one or more processors, cause the one or more processors to implement the training method for the hard landing risk prediction model provided in the first aspect of the invention and the hard landing risk prediction method provided in the second aspect.

[0016] This invention addresses the weakness of existing technologies in predicting hard landing risks by training a personalized hard landing risk prediction model for each pilot. In the early stages of model training, hard landing feature data strongly correlated with hard landings is directly selected from the pilot's complete historical flight and landing data. To enable the model to learn the pilot's operational patterns, trends, and habits, the hard landing feature data is augmented to obtain multi-dimensional augmented hard landing feature data. Furthermore, to enable the model to learn the pilot's flight habits and operational patterns across long, medium, and short time scales, the augmented hard landing feature data is categorized by time scale, constructing a training dataset for the initial hard landing risk prediction model corresponding to the pilot. This training dataset includes augmented hard landing feature data corresponding to each time scale and landing labels corresponding to the hard landing feature data. This dataset provides high-quality input for model feature learning and classification prediction.

[0017] Finally, the initial hard landing risk prediction model is trained using the training dataset to obtain an individualized hard landing risk prediction model corresponding to the pilot.

[0018] Because hard landing risk prediction models can learn key features highly correlated with hard landings and deeply fit the individualized operating habits of corresponding pilots, they possess the advantages of strong targeting, high prediction accuracy, and good adaptability. They have a high risk prediction capability, improving the effectiveness, targeting, adaptability, and practicality of risk prediction. In practical applications, as long as the flight and landing data of the flight to be predicted and the pilot information are obtained, the matching model can be invoked to predict the landing behavior of the flight. This allows for advance warning to the crew, intervention, and correction of landing operations, effectively reducing the risk of hard landings. The models are more practical and the prediction results are more reliable.

[0019] These features and advantages of the present invention will be disclosed in detail in the following specific embodiments and accompanying drawings. The preferred embodiments or means of the present invention will be shown in detail in conjunction with the accompanying drawings, but are not intended to limit the technical solutions of the present invention. In addition, each of these features, elements and components appearing in the following text and drawings is a plurality of, and different symbols or numbers are used for convenience of representation, but all represent parts with the same or similar construction or function. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings: Figure 1 A flowchart illustrating a training method for a hard landing risk prediction model provided in an embodiment of this application; Figure 2This is a flowchart illustrating a process for filtering re-landing feature data based on a pilot's historical flight and landing data, as provided in an embodiment of this application. Figure 3 This is a flowchart illustrating a process for filtering various flight landing parameters from historical flight landing data to generate re-landing spatial feature data, as provided in an embodiment of this application. Figure 4 This is a flowchart illustrating a process for filtering various flight landing parameters in historical flight landing data to generate relanding time feature data, as provided in an embodiment of this application. Figure 5 This is a flowchart illustrating a process for augmenting the hard landing feature data to generate augmented hard landing feature data, provided in an embodiment of this application. Figure 6 A schematic diagram of the model structure of a hard landing risk prediction model provided in an embodiment of this application; Figure 7 A flowchart illustrating a method for predicting the risk of a hard landing provided in an embodiment of this application; Figure 8 This application provides a schematic diagram of the structure of a system. Figure 9 This is a structural schematic diagram of an electronic device provided in an embodiment of this application.

[0021] Explanation of reference numerals in the attached figures Among them, 101 is the processor; 102 is the memory; 103 is the I / O interface; and 104 is the bus. 201. Filtering module; 202. Expansion module; 203. Generation module; 204. Training module; 205. Prediction module. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described are intended to explain the present invention and should not be construed as limiting the invention.

[0023] The terms "an embodiment," "example," or "trademark" used in this specification refer to a particular feature, structure, or characteristic described in connection with the embodiment itself that may be included in at least one embodiment disclosed in this invention. The phrase "in an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0024] In existing machine learning-based risk prediction methods, massive amounts of flight data are typically used as the foundation for model training, followed by binary classification models for feature learning and prediction. While large volumes of flight data can improve the generalization of the prediction model, these massive datasets contain a large amount of irrelevant data with weak correlation to hard landings. Inputting this irrelevant data into the model causes it to learn redundant and irrelevant features, thereby reducing the model's risk prediction accuracy.

[0025] Furthermore, the model's analysis of flight data is based on group statistics and generalized analysis. This means that a large amount of flight data over a specific period is input into the model without considering individual pilot operational characteristics (such as leveling deviations, aggressive stick pulls, energy management habits, etc.) that are strongly correlated with the probability of a hard landing. This makes it difficult for the model to capture individual operational habits and potential hard landing risks. Therefore, it suffers from low risk prediction probability and weak early warning capabilities.

[0026] In view of this, in order to improve the prediction capability and accuracy of hard landing risk, as a first aspect of the present invention, a training method for a hard landing risk prediction model is provided, such as... Figure 1 As shown, the training method includes: Step S110: Based on the pilot's historical flight and landing data, filter the hard landing feature data.

[0027] For example, it should be noted that historical flight landing data is recorded by the onboard Quick Access Recorder (QAR), including flight landing parameters corresponding to multiple flights, such as load, flight altitude, horizontal displacement, lateral track deviation, pitch angle, roll angle, yaw angle, airflow stability, runway approach altitude, landing wind direction, rate of descent, and vertical acceleration. Each flight landing parameter in the hard landing characteristic data is greater than or equal to a preset physical parameter threshold. The preset physical parameter threshold is used to determine whether the flight landing parameters exceed the specified threshold; if they do, there is a risk of hard landing. For example, vertical acceleration ≥ 2.0g. During the screening process, each flight landing parameter in the historical flight landing data is screened according to both time and spatial dimensions. Multiple flight landing parameters that exceed the specified threshold and are strongly correlated with hard landing behavior are screened as hard landing spatial characteristic data and hard landing temporal characteristic data, which together constitute the hard landing characteristic data.

[0028] Step S120: Expand the hard landing feature data to generate expanded hard landing feature data.

[0029] For example, feature engineering can be used to augment the data.

[0030] Step S130: Classify the expanded hard landing feature data according to time scale to generate a training dataset for the initial hard landing risk prediction model corresponding to the pilot; wherein, the training dataset includes expanded hard landing feature data corresponding to each time scale and landing labels corresponding to the hard landing feature data.

[0031] For example, the time scale can be the number of flights in the near term, medium term, and long term, as well as the flight and landing data, pilot information, and landing results for each flight.

[0032] Step S140: Train the initial hard landing risk prediction model based on the training dataset to generate a hard landing risk prediction model corresponding to the pilot.

[0033] For example, after iteratively training the model parameters of the initial hard landing risk prediction model based on the training dataset, a hard landing risk prediction model corresponding to the pilot is generated.

[0034] This invention addresses the weakness of existing technologies in predicting hard landing risks by training a personalized hard landing risk prediction model for each pilot. In the early stages of model training, hard landing feature data strongly correlated with hard landings is directly selected from the pilot's complete historical flight and landing data. To enable the model to learn the pilot's operational patterns, trends, and habits, the hard landing feature data is augmented to obtain multi-dimensional augmented hard landing feature data. Furthermore, to enable the model to learn the pilot's flight habits and operational patterns across long, medium, and short time scales, the augmented hard landing feature data is categorized by time scale, constructing a training dataset for the initial hard landing risk prediction model corresponding to the pilot. This training dataset includes augmented hard landing feature data corresponding to each time scale and landing labels corresponding to the hard landing feature data. This dataset provides high-quality input for model feature learning and classification prediction.

[0035] Finally, the initial hard landing risk prediction model is trained using the training dataset to obtain an individualized hard landing risk prediction model corresponding to the pilot.

[0036] Because hard landing risk prediction models can learn key features highly correlated with hard landings and deeply fit the individualized operating habits of corresponding pilots, they possess the advantages of strong targeting, high prediction accuracy, and good adaptability. They have a high risk prediction capability, improving the effectiveness, targeting, adaptability, and practicality of risk prediction. In practical applications, as long as the flight and landing data of the flight to be predicted and the pilot information are obtained, the matching model can be invoked to predict the landing behavior of the flight. This allows for advance warning to the crew, intervention, and correction of landing operations, effectively reducing the risk of hard landings. The models are more practical and the prediction results are more reliable.

[0037] In some embodiments, as an optional implementation of step S110, such as Figure 2 As shown, the process of filtering relanding feature data based on the pilot's historical flight and landing data includes: Step S111: Filter the various flight landing parameters of the historical flight landing data to generate re-landing spatial feature data.

[0038] Step S112: Filter the various flight landing parameters of the historical flight landing data to generate relanding time feature data.

[0039] For example, the historical flight landing data includes multiple flight landing parameters; each flight landing parameter in the re-landing feature data is greater than or equal to a preset physical parameter threshold; the re-landing feature data includes re-landing spatial feature data and re-landing temporal feature data. It should be noted that although the inputs to steps S111 and S112 are the same—each flight landing parameter from the historical flight landing data—their filtering objectives are different. The filtering objective of step S111 is to further filter out multiple flight landing parameters strongly correlated with re-landing behavior from a spatial dimension, forming re-landing spatial feature data; the filtering objective of step S112 is to further filter out multiple flight landing parameters strongly correlated with re-landing behavior from a temporal dimension, forming re-landing temporal feature data. In the actual filtering process, these two steps are independent of each other and do not affect each other; they can be executed in parallel, effectively improving the efficiency of data filtering.

[0040] In some embodiments, to accurately filter out spatial feature data of hard landing that are strongly correlated with hard landing behavior and improve the accuracy of the risk prediction model, as an optional implementation of S111, such as... Figure 3 As shown, the step of filtering the various flight and landing parameters of the historical flight and landing data to generate re-landing spatial feature data includes: Step S111a: Determine the contribution of each flight landing parameter to the smaller prediction error of the classification model based on the classification model.

[0041] For example, the contribution of the smaller prediction error is the contribution of the minimum prediction error. Taking LightGBM as a classification model as an example, the process of determining the contribution of each flight landing parameter in the historical flight landing data using LightGBM is given.

[0042] First, let's explain the principles of LightGBM. LightGBM is an open-source, lightweight decision tree ensemble model based on a gradient boosting framework (Light Gradient Boosting Machine). It's an additive model composed of multiple independent decision trees strung together. Each individual decision tree is a binary classification / regression tree, with a hierarchical structure of root node - intermediate splitting nodes - leaf nodes. Each node splits into left and right child nodes based on a single feature and the optimal threshold. In the entire decision tree ensemble model, each tree learns to compensate for the prediction errors of all preceding trees. Finally, the prediction results of all trees are weighted and summed to obtain the final output of the decision tree ensemble model.

[0043] Historical flight landing data includes not only various flight landing parameters but also landing results. Landing results are the landing outcomes (such as hard landings and normal landings) of multiple flights flown by the pilot during the approach and landing phase.

[0044] The LightGBM model takes flight and landing parameters from multiple flights operated by the same pilot as input, with the landing results of the multiple flights serving as labels. The loss function of the LightGBM model is log loss.

[0045] After receiving the samples and labels, the root node of the first decision tree in the LightGBM model first calculates the global initial prediction value (e.g., if the relanding rate in the samples is 10%, the initial prediction value is 0.1), and then calculates the initial log loss value of the first decision tree based on the global initial prediction value. The initial log loss value represents the initial prediction error of the model for the relanding label when there is no feature split. After that, the root node passes all sample data and the initial log loss value to the subsequent intermediate split nodes.

[0046] After receiving all sample data and the initial logarithmic loss value, the intermediate split node takes each flight landing parameter (such as load, airflow stability, runway approach altitude, landing wind direction, descent rate, etc.) as a candidate split feature. For each candidate split feature, it automatically traverses all possible numerical thresholds (such as the threshold for vertical acceleration being 1.8g, 2.0g, etc. according to preset physical parameter thresholds). According to the rules of single feature and single threshold, the sample is divided into a left child node sample set (such as vertical acceleration ≥ 2.0g) and a right child node sample set (such as vertical acceleration < 2.0g). Each combination of a single feature and a single threshold constitutes a split, and the split gain corresponding to a single split satisfies the following formula (1): (1) In equation (1) above, I is the sample set for each single feature; , These are the left and right child node sample sets after the split; , λ and γ are the first and second gradients of the sample, respectively; λ and γ are the regularization parameters, respectively; i is the selected feature (flight and landing parameters); The split gain represents the reduction in the loss function caused by each split. A larger split gain results in a greater reduction in the loss function with each split, a larger decrease in the loss function, and a smaller prediction error of the landing result. Conversely, a smaller split gain results in a smaller reduction in the loss function with each split, a weaker effect on reducing the loss function, and a smaller contribution to reducing prediction error.

[0047] According to formula (1) above, LightGBM selects the threshold for feature i in each round based on the maximum split gain before splitting nodes. In other words, the combination of the selected feature and threshold represents the initial split with the strongest ability to distinguish re-landing labels, and the corresponding flight and landing parameters are the features that contribute the most to reducing the prediction error. This is because only splitting according to the maximum classification gain can significantly reduce the model's loss function.

[0048] After splitting according to the optimal feature and threshold combination, two leaf nodes will be formed. The leaf node is the end point of the first tree and will not be split again. Each leaf node outputs the error fitting value (i.e., the correction amount to the initial prediction value) based on the label distribution of its own sample set; then, the overall prediction value of the model is updated according to the error fitting value, and the log loss value (i.e., prediction error) after the first tree split is determined; finally, the optimal splitting feature of the first tree (such as landing load), the split gain of this feature, and the prediction error after the split are recorded.

[0049] At this point, the first decision tree has completed its split, and the resulting optimal splitting feature and corresponding splitting gain (such as the splitting gain of the landing payload) are stored. =6.0), and the prediction error of the model after the first split is passed to the second decision tree as the fitting target of the second decision tree.

[0050] In the second tree and all subsequent decision trees, the same node splitting action as in the first tree is repeated, and the splitting gains corresponding to the same features are accumulated according to formula (2) to obtain the contribution of the splitting gain corresponding to the feature in the entire LightGBM decision tree ensemble model (the contribution of the minimum prediction error). Each decision tree will successively fit the prediction errors of all previous trees and take minimizing the current remaining prediction error as the fitting objective.

[0051]

[0052] In the above formula (2), j is the feature (i.e. the selected flight and landing parameters). The cumulative splitting gain (i.e., the splitting gain contribution) of feature j. This represents the splitting gain after splitting feature j at this node. This indicates that all decision trees in the entire ensemble model are traversed to prevent misjudgment of feature contributions. It is the summation of the split gains for all nodes that use feature j as a split node.

[0053] Finally, after multiple iterations, the log loss value of the LightGBM model gradually converged. At this point, most of the flight landing parameters (features) that are discriminative and relevant to the relanding label have been mined, and the split gain contribution of each feature has also been recorded. The split gain contributions of each feature are sorted in descending order. The higher the split gain contribution, the greater the overall contribution of that feature to reducing the relanding label prediction error in the entire iterative splitting of the LightGBM model, that is, the stronger its correlation with relanding behavior.

[0054] Step S111b: Sort each flight landing parameter according to the contribution degree, and select the first to Nth flight landing parameters after sorting; generate relanding spatial feature data based on the first to Nth flight landing parameters after sorting, wherein N is greater than 1 and the contribution degree of the Nth flight landing parameter is greater than or equal to the preset contribution degree.

[0055] For example, continuing with the split gain contribution of each feature determined by the LightGBM model, the specific steps for sorting are given. The features are sorted in descending order of their split gain contribution, and then the top 23 flight landing parameters strongly correlated with hard landing are selected in descending order, as shown in Table 1 below. These selected flight landing parameters constitute the hard landing spatial feature data. Table 1 shows the selected hard landing spatial feature data.

[0056] Table 1

[0057] The aforementioned screening method is essentially a "false prediction, true screening" approach utilizing existing classification models. It uses only the classification task of the existing model as a vehicle, leveraging the model's iterative process (aiming for minimum prediction error) to quantify the influence of each flight landing parameter on the prediction result. The classification model accurately measures the importance (contribution to minimizing prediction error) of each flight landing parameter (feature). A higher contribution indicates a stronger correlation between the feature and the hard landing behavior, and greater value for the hard landing risk prediction result. Through this method, useful features for hard landing prediction results can be accurately screened, while useless flight landing parameters are discarded, providing high-quality input for subsequent prediction models and improving their accuracy.

[0058] It is important to explain the beneficial effects of using the LightGBM model to filter historical flight landing data: As is well known to those skilled in the art, airborne QAR records a large number of flight parameters. The data filtering method described in this application can automatically reduce the dimensionality of the original high-dimensional data while eliminating redundant and low-information-content parameters, thus efficiently and effectively identifying features strongly correlated with hard landings. This not only provides high-quality input for subsequent hard landing risk prediction models but also reduces the training cost and speed of these models, effectively ensuring the accuracy of subsequent prediction models. Simultaneously, it solves the problems of low efficiency and easy omission of potentially related features in manual screening.

[0059] Unlike traditional linear data screening and preprocessing methods, this application leverages the nonlinear fitting capabilities of decision trees to effectively capture the complex nonlinear correlation between flight landing parameters and hard landing risk, accurately reflecting the inherent complex patterns in the data. Furthermore, by training the LightGBM model multiple times and calibrating its results, the quantification of feature importance becomes more precise, and the screening results become more stable.

[0060] It is also important to note that this application differs from the conventional use of LightGBM. While this application utilizes the LightGBM model to perform binary classification predictions (re-landing, normal landing) on ​​historical flight landing data, it does not use the prediction results as the final re-landing risk prediction result. Instead, it uses the LightGBM model's prediction results to quantify the contribution of each flight landing parameter in the historical flight landing data to the re-landing label prediction, filtering out multiple flight landing parameters strongly correlated with the occurrence of a re-landing event, thus obtaining re-landing spatial feature data. In other words, the LightGBM model prediction is a kind of "pseudo-prediction, true filtering." It uses the prediction task as a vehicle, leveraging the ability to quantify the importance of features during the prediction process to filter out features useful for the re-landing prediction result; the prediction result itself has no practical business significance.

[0061] More importantly, the input data in this application is the historical flight and landing data of a single pilot. It specifically mines hard landing risk correlation parameters unique to the pilot, rather than mining general patterns for a group. This allows the subsequent hard landing risk prediction model to accurately learn the pilot's operating habits and significantly improve the accuracy of hard landing risk prediction for the flight operated by that pilot.

[0062] To further improve the input data quality and predictive capability of the hard landing risk prediction model, in some embodiments, this application also provides another independent time-series index screening process, aimed at filtering hard landing time feature data that is strongly correlated with time-series operations from historical flight landing data. As one implementation of step S112, such as... Figure 4 As shown, the step of filtering the various flight and landing parameters of the historical flight and landing data to generate relanding time feature data includes: Step S112a: Determine the target entropy of each flight landing parameter within the set flight landing time period based on the number of times each flight landing parameter appears within the set flight landing time period.

[0063] For example, the following is a specific implementation of filtering dynamic time-series flight landing parameters that are strongly correlated with hard landing and have high information content from the massive flight landing data of airborne QAR.

[0064] First, obtain time series samples. The time periods within which the flight stage is 6 <= flight_stage <= 12 and the flight altitude is 0.5 <= radio_attitude <= 50 are defined as the flight and landing time periods. Then, obtain the various flight and landing parameters for each of these time periods.

[0065] It's important to clarify the concept of flight stages. "Flight_stage" is a digital code for the flight stage of a civil aircraft, and 6 <= flight_stage <= 12 represents the standardized coding range within the civil aviation industry for the core low-altitude phase from approach to landing. The corresponding operational procedures for this flight stage include five-way approach preparation, five-way approach, leveling off, and approaching at 0.5 feet above ground. This phase involves high-frequency pilot maneuvers and is also a high-risk period for hard landings. Limiting the stage to 6 <= flight_stage <= 12 allows for the extraction of data unrelated to hard landings, retaining only the critical operational phases before landing. This allows subsequent entropy calculations and the selection of hard landing time feature data to focus more on pilot operation time-series data strongly correlated with hard landing risks, improving the accuracy of feature selection and the training efficiency of subsequent models.

[0066] `radio_attitude` represents the flight altitude in feet. `0.5 ≤ radio_attitude ≤ 50` indicates the approach and landing phase within the first 50 feet of landing. This phase is also the period of high-frequency pilot operations, including controlling the rate of descent, stick inputs, and pitch adjustments. These operations are crucial in determining whether a hard landing will occur; therefore, it is necessary to define the time period by flight phase and altitude, and extract all time-related flight and landing parameters within that time period.

[0067] Secondly, taking the descent rate as a time-related flight and landing parameter and the target entropy calculation as a normalized permutation entropy as an example, the entropy calculation in step S112a is as follows: First, the descent rate within the set flight and landing time period is divided into several data windows of equal length. For example, if there are 1000 millimeter-level data points of continuous descent rate data, after dividing them into 200 data windows of equal length, each data window contains 5 data points. The second step is to determine the arrangement pattern of the data windows based on the magnitude of the multiple decrease rates of the data windows. For example, if the decrease rate of the first data window continuously decreases, and is successively 500ft / min, 450ft / min, 400ft / min, 350ft / min, and 300ft / min, then its corresponding arrangement pattern is "54321"; if the decrease rate of the second data window fluctuates, and is successively 400ft / min, 500ft / min, 450ft / min, 350ft / min, and 400ft / min, then its corresponding arrangement pattern is "35423"; if the decrease rate of the third data window continuously increases, and is successively 300ft / min, 350ft / min, 400ft / min, 450ft / min, and 500ft / min, then its corresponding arrangement pattern is "12345". The third step is to determine the probability of each arrangement pattern appearing in all data windows according to formula (3).

[0068] (3) In formula (3), π represents a permutation pattern; d represents the embedding dimension; for example, if the "54321" permutation pattern appears 20 times in 200 data windows, the probability corresponding to this pattern is 20 / 200=0.1.

[0069] The fourth step is to substitute the probability of the permutation pattern determined by formula (3) into formula (4) to obtain the normalized permutation entropy.

[0070] (4) In formula (4), To normalize the permutation entropy, the meanings of π and d are consistent with those in formula (3). As obtained from formula (4) The closer the value is to 1, the higher the temporal information content of the flight landing parameter (such as the descent rate in this embodiment), and the stronger its correlation with the risk of a hard landing. If the data of a certain flight landing parameter during the flight landing period remains almost unchanged across all data windows, that is, if the arrangement pattern of all data windows is "11111" (100% repetition), then the calculated value... A value of 0 indicates that the parameter has zero information content and is of little analytical value.

[0071] Based on the above steps, the entropy (normalized permutation entropy) of each flight and landing parameter within the set flight and landing time period can be obtained.

[0072] Step S112b: Sort each flight landing parameter based on the target entropy, and select the first to Mth flight landing parameters after sorting; generate relanding time feature data based on the first to Mth flight landing parameters, wherein M is greater than 1 and the entropy of the Mth flight landing parameter is greater than or equal to the preset entropy.

[0073] For example, multiple normalized permutation entropies are filtered in descending order to select several flight and landing parameters that are strongly correlated with the hard landing, as shown in Table 2 below. The selected multiple (first 13) flight and landing parameters constitute the hard landing time feature data. Table 2 shows the selected hard landing time feature data.

[0074] Table 2

[0075] After obtaining the spatial and temporal features of the re-landing data using the aforementioned screening method, to enrich their information dimensions, enhance the model's feature representation capabilities, and improve the accuracy of its predictions, the selected data are augmented to obtain augmented spatial and temporal re-landing feature data. These two data sets together constitute the augmented re-landing feature data. Since the data screening process is independent, the augmentation processes are also independent and parallel, without mutual interference.

[0076] In some embodiments, as an optional implementation of step S120, such as Figure 5 As shown, the step of augmenting the hard landing feature data to generate augmented hard landing feature data includes: Step S121: Based on feature engineering, determine multiple statistical dimensions of the hard landing spatial feature data, and generate derived hard landing spatial feature data corresponding to each statistical dimension.

[0077] For example, feature engineering includes centralized distribution statistics, discreteness statistics, time-domain trend statistics, and extreme value statistics. The corresponding statistical dimensions are centralized operations (such as mean, median, mode), discrete operations (such as standard deviation, variance), time-domain operations (such as difference, integral), and extreme value operations (maximum value, minimum value).

[0078] Step S122: Based on the hard landing spatial feature data and the corresponding derived hard landing spatial feature data, generate expanded hard landing spatial feature data; For example, it should be noted that steps S121-S122 above are for expanding the heavy landing spatial feature data. Taking the landing load in the heavy landing spatial feature data as an example, the landing load is a value, a one-dimensional feature. By calculating the mean, variance, maximum value, and minimum value of the landing load of the recent 30 flights, data corresponding to four new statistical dimensions are added. The mean, variance, maximum value, and minimum value of the landing load are the derived heavy landing spatial feature data. The four derived heavy landing spatial feature data and the original landing load are used together as the expanded heavy landing spatial feature data.

[0079] Similarly, by calculating the maximum and minimum putts below 50ft and below 30ft respectively, we can add derivative feature data for the extreme value dimension.

[0080] Step S123: Based on feature engineering, determine multiple statistical dimensions of the hard landing time feature data, and generate derived hard landing time feature data corresponding to each statistical dimension.

[0081] For example, steps S123-S124 are for expanding the re-landing time feature data; they can be executed in parallel with steps S121-S122, which can effectively improve the efficiency of data expansion.

[0082] Step S124: Generate extended hard landing time feature data based on the hard landing time feature data and the corresponding derived hard landing time feature data.

[0083] For example, the hard landing time feature data and the corresponding derived hard landing time feature data are used as extended hard landing time feature data. The extended hard landing feature data includes extended hard landing spatial feature data and extended hard landing time feature data. Taking the nose-up stick amount in the hard landing time feature data as an example, the nose-up stick amount is a value, which is a one-dimensional feature. By obtaining the parameter values ​​of the nose-up stick amount of the most recent 50 flights and calculating the difference of the nose-up stick amount of adjacent flights, the changing trend of stick amount operation can be obtained, thus adding a new dimension of feature data.

[0084] Furthermore, a time-series curve is constructed with flight number as the x-axis and the corresponding nose-up stick movement as the y-axis. Integrating the nose-up stick movement parameter value sequence of the recent 50 flights yields the area under the curve (AUC), which reflects the overall operational amplitude of the nose-up stick movement. This adds another dimension of feature data. The difference and integration operations essentially obtain the dynamic variation patterns of the hard landing time feature data, using these patterns as derived hard landing time feature data.

[0085] Unlike traditional data augmentation techniques, this application expands the data dimension of features. It calculates derived feature data through feature engineering, adding multiple feature indicators to the original feature data to obtain a high-dimensional feature set. Traditional data augmentation expands the sample set, such as by using image flipping or text synonym replacement to increase the number of training samples and alleviate model overfitting. Applying traditional data augmentation to the civil aviation QAR flight data application domain, the corresponding method is to increase the number of flight samples.

[0086] The data augmentation method in this application avoids simply repeating and stacking single parameter values ​​from a single flight. Instead, it incorporates the pilot's long-term operational patterns, trends, and habits into the feature dimension, making the model input more representative and richer in information. This allows the model to learn the personalized operational patterns of pilots that are closely related to hard landings, effectively improving the model's predictive capabilities.

[0087] To improve the ability and accuracy of predicting hard landing risks, this application also constructs a hard landing risk prediction model. The structure of the hard landing risk prediction model is described in detail below.

[0088] The initial hard landing risk prediction model includes: a cascaded feature extraction module and a feature fusion classification module, wherein the feature extraction module includes multiple parallel feature extraction sub-modules corresponding to time scales; The feature extraction submodule is used to extract spatiotemporal features of the augmented relanding feature data corresponding to the time scale. The feature fusion classification module is used to fuse the spatiotemporal features corresponding to each extracted time scale and output the classification result. The classification results include hard landing and normal landing.

[0089] Each feature extraction submodule includes cascaded one-hot encoding, a spatiotemporal encoder, merging, and an attention mechanism. First, one-hot encoding is performed on the augmented relanding temporal feature data and the augmented relanding spatial feature data, generating augmented relanding temporal feature vectors and augmented relanding spatial feature vectors, respectively. Second, Reshape is used to reconstruct the dimensions of the augmented relanding temporal feature vectors and the augmented relanding spatial feature vectors, generating fixed-dimensional augmented relanding temporal feature structures and augmented relanding spatial feature structures. Next, a temporal value encoder (LSTM) based on the spatiotemporal encoder extracts features from the augmented relanding temporal feature structures, generating temporal features; a measurement value encoder (CNN) based on the spatiotemporal encoder extracts features from the augmented relanding spatial feature structures, generating spatial features. Then, the temporal features and the spatial features are merged to generate merged features. Finally, the merged features are weighted and aggregated using an attention mechanism to obtain spatiotemporal features; wherein, the spatiotemporal features include contextual semantic features corresponding to the time scale. This completes the steps of the feature extraction submodule to extract the spatiotemporal features of the augmented hard landing feature data. Each feature extraction submodule corresponds to a different time scale (i.e., a different number of flights are selected), so multiple feature extraction submodules can extract the spatiotemporal features of the augmented hard landing feature data corresponding to the time scale.

[0090] It should be noted that different time scales can reflect a pilot's historical performance. For example, monitoring physiological and emotional states and sudden control deviations in the recent 30 flights; assessing changes in technical stability and dynamic trends in skills in the medium term using 100 flights; and establishing an individual operational baseline using 200 flights in the long term. Through different feature extraction sub-modules, the model can ultimately learn the inherent control habits and styles specific to a particular pilot.

[0091] The feature fusion and classification module includes a cascaded feature fusion layer, a fully connected layer, a Dropout layer, and a binary classifier. First, the spatiotemporal features corresponding to each time scale are input to the feature fusion layer for feature fusion, resulting in multi-scale fused features. Second, the multi-scale fused features are input to the fully connected layer to obtain feature vectors for the binary classifier's classification space. Next, these feature vectors are input to the Dropout layer for random deactivation, resulting in regularized feature vectors to prevent overfitting. Finally, the regularized feature vectors are input to the binary classifier to obtain the classification result. This completes the fusion and classification of spatiotemporal features corresponding to each time scale.

[0092] For example, Figure 6 This is a schematic diagram of a hard landing risk prediction model of the present invention. The model input consists of hard landing time feature data (TS Input) and hard landing spatial feature data (Meas Input) at different time scales. The dimensions of TS Input are [BS, N, 12, 13], where 12 is the number of channels and 13 is the number of features corresponding to each channel (i.e., the first 13 flight and landing parameters of the selected hard landing time feature data). The dimensions of Meas Input are [BS, N, 1, 25], where 1 is the number of channels and 25 is the number of features corresponding to each channel (i.e., the first 25 flight and landing parameters of the selected hard landing spatial feature data); BS is the Batch_Size of the input model, i.e., the batch number; N is the sequence length, i.e., the number of consecutive time steps contained in each input sample. The model supports three configurations: N=30, 100, and 200 flights, corresponding to short, medium, and long flight operation segments, respectively, ensuring that the model can capture operational habit features at different time scales.

[0093] The TS Input and Meas Input are one-hot encoded and reshaped to achieve dimensionality adaptation. One-hot encoding is used to convert discrete operational features into continuous vectors that the model can process, while reshape reconstructs high-dimensional tensors into flat structures suitable for subsequent encoder inputs. This eliminates the dimensionality inconsistency problem caused by different sequence lengths (30 / 100 / 200) and ensures that the model can be compatible with the analysis of flight segments of different durations.

[0094] Next, the dual encoder feature extraction module is entered. The temporal value encoder (LSTM) captures the long-term dependencies of the pilot's operations through a gating mechanism and extracts the temporal dynamic features that reflect the operating style. The measurement value encoder (CNN) extracts the spatial local patterns in the Meas Input through the sliding of the convolution kernel and mines the state response patterns that are strongly associated with the operating habits. The features of the two modalities are concatenated and merged to form a unified feature representation with a dimension of (BS*N, 128), realizing the deep fusion of operating behavior and state response.

[0095] Next, the feature combination performs nonlinear transformation and enhancement on the merged features through a fully connected layer, further improving the discriminative power and expressive power of the features; shape rollback restores the flattened feature sequence to a batch sequence structure of (BS, N, 128), restores the temporal order attribute of the flight segment, adapts to the input requirements of the attention mechanism for sequence data, and lays the foundation for subsequent focusing on key operation segments.

[0096] Subsequently, the attention mechanism pooling learns the weight distribution of features at each time step, automatically focusing on the key operational segments most representative of the pilot's habits, generating contextual semantic features of dimension (BS, 128), and outputting attention weights (BS, N). This not only achieves efficient feature aggregation but also provides interpretable basis for model decisions, facilitating the analysis of the operational focus of the model. In the figure, APSW represents Attention Pooling Specific Weights; CSF represents Contextual Semantic Features. The feature dimension obtained by feature fusion is (BS, 384).

[0097] Finally, the classification decision module deeply integrates the contextual semantic features through a feature fusion layer, strengthening the correlation between features of different dimensions. Then, a fully connected layer maps the high-dimensional features to the classification space, ultimately outputting a binary classification Logits (whether a relanding occurred) with a dimension of (BS,2), used to determine whether the current operation conforms to the pilot's long-established baseline habits. The fully connected layer has 256 units, the activation layer uses reLU, and the output classification result is Final Logits (BS,2) binary classification.

[0098] As a second aspect of the present invention, a method for predicting the risk of a hard landing is provided, such as... Figure 7 As shown. The heavy landing risk prediction method includes: Step S210: Obtain the pilot information and flight and landing data for the current flight.

[0099] Step S220: Filter the flight landing data to generate hard landing feature data.

[0100] Examples include the screening embodiments disclosed above for step S110.

[0101] Step S230: Expand the hard landing feature data to generate expanded hard landing feature data.

[0102] Examples include several extended embodiments of step S120 as described above.

[0103] Step S240: Input the expanded hard landing feature data into the hard landing risk prediction model trained by the training method of the first aspect of the present invention to generate the hard landing risk prediction result of the current flight.

[0104] For example, the input obtained in step S230 is input into a trained hard landing risk prediction model matched with the pilot to obtain the hard landing risk prediction result for the current flight. Based on the prediction result, the crew can be alerted in advance to intervene and correct the landing operation, thereby effectively reducing the hard landing risk. This method is more practical and the prediction effect is more reliable.

[0105] As a third aspect of the invention, a hard landing risk prediction system is also provided, such as... Figure 8 As shown, it includes: The filtering module 201 is used to filter out hard landing feature data from the pilot's historical flight landing data; The expansion module 202 is used to expand the hard landing feature data to generate expanded hard landing feature data; The generation module 203 is used to classify the expanded hard landing feature data according to time scales and generate a training dataset for the initial hard landing risk prediction model corresponding to the pilot; wherein, the training dataset includes expanded hard landing feature data corresponding to each time scale and landing labels corresponding to the hard landing feature data; Training module 204 is used to train the initial hard landing risk prediction model using the training dataset to generate a hard landing risk prediction model corresponding to the pilot. The prediction module 205 is used to input the expanded hard landing feature data into the hard landing risk prediction model corresponding to the acquired pilot information, and generate the hard landing risk prediction result of the current flight, wherein the hard landing risk prediction model is a model generated by the training module.

[0106] Furthermore, as a fourth aspect of the present invention, an electronic device is provided, such as... Figure 9 As shown, it includes: One or more processors 101; The memory 102 stores one or more computer programs that, when executed by the one or more processors 101, cause the one or more processors 101 to implement the training method for the hard landing risk prediction model provided in the first aspect of the present invention and the hard landing risk prediction method provided in the second aspect.

[0107] The tool may also include one or more I / O interfaces 103 connected between the processor 101 and the memory 102, configured to enable information interaction between the processor 101 and the memory 102.

[0108] The processor 101 is a device with data processing capabilities, including but not limited to a central processing unit 101 (CPU); the first memory 102 is a device with data storage capabilities, including but not limited to random access memory 102 (RAM, more specifically SDRAM, DDR, etc.), read-only memory 102 (ROM), electrically erasable programmable read-only memory 102 (EEPROM), and flash memory (FLASH); the I / O interface 103 (read-write interface) is connected between the processor 101 and the memory 102, enabling information interaction between the processor 101 and the memory 102, including but not limited to a data bus 104 (Bus).

[0109] In some embodiments, the processor 101, memory 102, and I / O interface 103 are interconnected via bus 104, and thus connected to other components of the computing device.

[0110] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art should understand that the present invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments above. Any modifications that do not depart from the functional and structural principles of the present invention will be included within the scope of the claims.

Claims

1. A training method for a heavy landing risk prediction model, characterized in that, include: Based on the pilot's historical flight and landing data, select re-landing characteristic data; The hard landing feature data is augmented to generate augmented hard landing feature data; The expanded hard landing feature data is classified according to time scale to generate a training dataset for the initial hard landing risk prediction model corresponding to the pilot. The training dataset includes the augmented relanding feature data corresponding to each time scale and the landing labels corresponding to the relanding feature data; The initial hard landing risk prediction model is trained based on the training dataset to generate a hard landing risk prediction model corresponding to the pilot.

2. The training method according to claim 1, characterized in that, The process of filtering relanding characteristic data based on pilots' historical flight and landing data includes: By filtering the various flight and landing parameters of the historical flight and landing data, re-landing spatial feature data is generated. By filtering the various flight and landing parameters of the historical flight and landing data, re-landing time feature data is generated. The historical flight and landing data includes multiple flight and landing parameters; Each flight landing parameter of the hard landing feature data is greater than or equal to a preset physical parameter threshold; the hard landing feature data includes hard landing spatial feature data and hard landing temporal feature data.

3. The training method according to claim 2, characterized in that, The process of filtering the historical flight and landing data for each flight and landing parameter to generate re-landing spatial feature data includes: The contribution of each flight landing parameter to the smaller prediction error of the classification model is determined based on the classification model. The flight and landing parameters are sorted according to their contribution, and the first to Nth sorted flight and landing parameters are selected. Based on the sorted first to Nth flight landing parameters, re-landing spatial feature data is generated, wherein N is greater than 1 and the contribution of the Nth flight landing parameter is greater than or equal to a preset contribution.

4. The training method according to claim 2, characterized in that, The process of filtering the historical flight and landing data for each flight and landing parameter to generate relanding time feature data includes: The target entropy of each flight landing parameter within the set flight landing time period is determined based on the number of times each flight landing parameter appears within the set flight landing time period. Based on the target entropy, sort the flight and landing parameters, and select the first to Mth sorted flight and landing parameters; Based on the first to M flight landing parameters, re-landing time characteristic data is generated; Where M is greater than 1 and the target entropy of the Mth flight landing parameter is greater than or equal to the preset entropy.

5. The training method according to any one of claims 1 to 4, characterized in that, The step of augmenting the hard landing feature data to generate augmented hard landing feature data includes: Based on feature engineering, multiple statistical dimensions of the hard landing spatial feature data are determined, and derived hard landing spatial feature data corresponding to each statistical dimension are generated. Based on the hard landing spatial feature data and the corresponding derived hard landing spatial feature data, expanded hard landing spatial feature data is generated; Based on feature engineering, multiple statistical dimensions of the hard landing time feature data are determined, and derived hard landing time feature data corresponding to each statistical dimension are generated. Based on the hard landing time feature data and the corresponding derived hard landing time feature data, extended hard landing time feature data is generated; The expanded hard landing feature data includes expanded hard landing spatial feature data and expanded hard landing temporal feature data.

6. The training method according to any one of claims 1 to 4, characterized in that, The initial hard landing risk prediction model includes: a cascaded feature extraction module and a feature fusion classification module. The feature extraction module includes multiple parallel feature extraction sub-modules corresponding to the time scale; The feature extraction submodule is used to extract the spatiotemporal features of the augmented re-landing feature data corresponding to the time scale; The feature fusion classification module is used to fuse the spatiotemporal features corresponding to each extracted time scale and output the classification result. The classification results include hard landing and normal landing.

7. The training method according to claim 6, characterized in that, The spatiotemporal features extracted from the augmented relanding feature data corresponding to the time scale include: One-hot encoding is performed on the augmented hard landing time feature data and the augmented hard landing spatial feature data respectively to generate augmented hard landing time feature vector and augmented hard landing spatial feature vector; The dimensions of the extended hard landing time feature vector and the extended hard landing space feature vector are reconstructed respectively to generate a fixed-dimensional extended hard landing time feature structure and an extended hard landing space feature structure. A time-series value encoder based on a spatiotemporal encoder extracts features from the augmented relanding time feature structure to generate time-series features; The measurement encoder based on the spatiotemporal encoder extracts features from the extended hard landing spatial feature structure to generate spatial features; The temporal features and spatial features are merged to generate merged features; The merged features are weighted and aggregated using an attention mechanism to generate spatiotemporal features. The spatiotemporal features include: contextual semantic features corresponding to the time scale.

8. A method for predicting the risk of a hard landing, characterized in that, include: Obtain pilot information and flight / landing data for the current flight; The flight and landing data are filtered to generate hard landing feature data; Expand the hard landing feature data to generate expanded hard landing feature data; The expanded hard landing feature data is input into the hard landing risk prediction model to generate the hard landing risk prediction result for the current flight; wherein, the hard landing risk prediction model is a model trained by a training method for a hard landing risk prediction model provided according to any one of claims 1 to 7.

9. A hard landing risk prediction system, characterized in that, include: The filtering module is used to filter out hard landing characteristic data from the pilot's historical flight landing data; An expansion module is used to expand the hard landing feature data to generate expanded hard landing feature data. The generation module is used to classify the expanded hard landing feature data according to time scales and generate a training dataset for the initial hard landing risk prediction model corresponding to the pilot; wherein, the training dataset includes expanded hard landing feature data corresponding to each time scale and landing labels corresponding to the hard landing feature data; The training module is used to train the initial hard landing risk prediction model using the training dataset to generate a hard landing risk prediction model corresponding to the pilot. The prediction module is used to input the expanded hard landing feature data into the hard landing risk prediction model corresponding to the acquired pilot information, and generate the hard landing risk prediction result for the current flight, wherein the hard landing risk prediction model is a model generated by the training module.

10. An electronic device, characterized in that, include: One or more processors; A memory having stored thereon one or more computer programs that, when executed by the one or more processors, cause the one or more processors to implement the method according to any one of claims 1 to 8.

Images