Intelligent assessment method for takeoff and landing stability based on simulator QAR data

Abstract

This invention belongs to the field of flight simulator takeoff and landing stability assessment technology, specifically involving an intelligent assessment method for takeoff and landing stability based on simulator QAR data. It aims to solve problems that hinder training efficiency and pose potential risks to flight safety. This invention acquires and decodes historical QAR data, employs a hierarchical adaptive sampling strategy to extract observation point parameters, and merges multi-stage data by combining identifier fields. It calculates the statistical values ​​of parameters at each observation point as reference values, uses an improved Lilliefors test to determine data distribution characteristics, and adaptively selects either a normal distribution model or a quantile group comparison model for stability assessment. It also introduces a machine learning-based parameter sensitivity score to dynamically adjust weights. Furthermore, it constructs a vector autoregression model and a Granger causality directed graph to achieve temporal tracing of parameter anomaly propagation paths. This invention effectively improves training efficiency and flight safety levels.

Description

Technical Field

[0001] This invention belongs to the field of simulator takeoff and landing stability assessment technology, and also relates to auxiliary decision-making applications for computer and auxiliary equipment repair. Specifically, it relates to a method and system for intelligent assessment of takeoff and landing stability based on simulator QAR data. Background Technology

[0002] In the field of flight training, traditional evaluation methods have long relied on the subjective observation and experience of instructors or inspectors. Evaluators must observe trainees' operational behavior in real time during training and manually score them according to the flight syllabus and scoring guidelines. This method is heavily influenced by the evaluator's personal experience, attention span, and fatigue, making it difficult to guarantee objectivity and consistency. Furthermore, it lacks the ability to conduct detailed and quantitative data analysis of the entire flight process. This approach completely severs the connection between training evaluation and the status monitoring of the flight simulator's computers and auxiliary equipment (such as the main control computer, control system modules, sensors, data transmission units, and control chips). It cannot use parameter feedback during training to infer the equipment's operational status, leading to long-term reliance on periodic shutdowns for manual inspection of simulator equipment.

[0003] With the popularization of flight data recording technology and the upgrading of computer data processing capabilities, automated evaluation methods based on rapid access to recorder data and intelligent computer analysis have gradually developed. Early methods could only achieve basic data statistics and presentation, limited by algorithm accuracy and computational efficiency, making it difficult to support in-depth analysis of complex flight phases. With breakthroughs in machine learning, time-series data analysis, and causal inference technologies, the intelligence level of evaluation systems has significantly improved, enabling dynamic correlation analysis of multiple parameters and anomaly identification. For example, the patent "Method for Evaluating Flight Approach and Landing Stability Based on QAR Data" (application number: 202110126329.2) proposes a systematic solution that, through the collaboration of modules such as data import, cleaning, parameter selection, model calculation, and visualization, achieves automated evaluation of the approach and landing process, improving evaluation efficiency and standardization to a certain extent. Currently, as flight simulator computer systems evolve towards higher computing power and integration, and with the intelligent upgrading of auxiliary equipment, combining QAR data evaluation with equipment status monitoring has become a core direction for supporting improved training efficiency and optimized equipment operation and maintenance.

[0004] However, existing technologies of this kind still have significant functional limitations when addressing the actual needs of pilot training. First, their analytical granularity remains coarse, focusing primarily on macroscopic evaluations of the final landing result, lacking continuous and dynamic analysis of each flight phase, such as takeoff, approach, leveling off, and touchdown, making it difficult to accurately pinpoint the specific stage at which a problem occurs. Second, while the system can present parameter trends, it has not established a dynamic correlation model between key flight parameters (such as pitch angle, rate of descent, airspeed, and throttle position) and the operating status of the computer and auxiliary equipment. Therefore, it cannot accurately trace the root cause parameters when landing deviations occur, nor can it reveal the propagation path of anomalies between different parameters, nor can it determine the type of equipment failure based on parameter anomaly characteristics (e.g., misjudging sensor data drift as pilot error). Furthermore, due to the lack of in-depth temporal analysis and causal inference capabilities, the conclusions output by existing technologies are often abstract, making it difficult to generate specific and actionable suggestions such as "improper pitch angle control at 50 feet led to an excessive touchdown rate," and it also fails to provide effective data support for simulator equipment repair.

[0005] Pilot training still lacks data-driven solutions that can provide accurate, traceable, and actionable feedback, which not only restricts the improvement of training efficiency but also poses hidden dangers to flight safety. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies—namely, that current aircraft landing process analysis techniques can only obtain final result assessments and lack data solutions that can provide accurate, traceable, and actionable feedback—which restricts training efficiency and poses potential risks to flight safety, this invention, in its first aspect, proposes an intelligent assessment method for takeoff and landing stability based on simulator QAR data. This method is used to determine the pilot's flight stability at various observation points, thereby obtaining stage stability. The method includes the following steps:

[0007] S1. Obtain and decode the historical QAR parameters of all pilots during training on the flight simulator to obtain QAR parameters in CSV format, which are used as the first data. Extract the first data corresponding to different flight states and merge the first data of the same flight segment in different flight stages according to the preset combination identifier field to obtain the merged data.

[0008] S2, Based on the merged data, calculate the statistical values ​​of the QAR parameters at different observation points in each flight phase, and determine the reference values ​​of the QAR parameters at each observation point;

[0009] S3, obtain the QAR parameters of the pilot training at different observation points for whom the flight stability is to be determined, as the second data; based on the second data, use the improved Lilliefers test method to determine whether the data distribution of the observation points follows a normal distribution; if yes, then determine the pilot's flight stability at each observation point based on the normal distribution; if no, then determine the pilot's flight stability at each observation point based on the position of the second data in the overall group distribution interval.

[0010] S4, Statistically analyze the flight stability at each observation point, and calculate the pilot's initial stage stability by combining the preset continuity threshold and proportion threshold.

[0011] S5. When an unstable observation point is identified, a causal network reflecting the dynamic influence relationship between parameters is constructed based on the QAR parameter time series of this flight. Based on the causal network, the abnormal propagation path leading to the instability is traced in reverse, and an abnormality source description is generated.

[0012] S6. Based on the preliminary stage stability and the anomaly tracing explanation, determine the final stage stability risk level.

[0013] The beneficial effects of this invention are:

[0014] By identifying the flight status, observation point parameters are selectively extracted from QAR data and decoded to form a flight segment CSV file. Using a combination of identification fields such as equipment number, event number, and training number, the flight parameters of the four stages of takeoff and landing are efficiently merged. Missing data is marked as null values, and a hierarchical adaptive sampling strategy is introduced. In the low-altitude drift stage, a dual-threshold sampling mechanism of time and event is adopted, which effectively solves the problems of easy omission and inaccurate data collection in key stages, ensuring data integrity and standardization, and laying a solid data foundation for subsequent analysis.

[0015] In the parameter analysis phase, the mean, standard deviation, and quantile statistical indicators of parameters at each observation point are calculated, reference values ​​are determined and stored, and the takeoff and landing phases are distinguished. An improved Lilliefors test (which incorporates quantile anomaly detection preprocessing) is used to calibrate the normal / non-normal distribution characteristics and parameter weights of the parameters. This allows for the intelligent and robust selection of either a parameter model based on normal distribution or a population comparison model based on quantiles for evaluation. Furthermore, a parameter sensitivity scoring and dynamic weight adjustment mechanism based on machine learning is introduced, which achieves non-linear enhancement of the evaluation weights of key control parameters, significantly improving the accuracy, flexibility, and intelligence of the analysis.

[0016] In the stability assessment stage, the stability of a single point is determined by comparing the parameters of the observation point with the selected reference distribution interval. The stability of the preliminary stage is calculated by combining the continuity threshold and the proportion threshold. This eliminates the subjectivity of traditional manual assessment and judges the stability status of each stage of flight in a quantitative and objective way. By introducing the vector autoregression (VAR) model and Granger causality test, a causal directed graph between parameters is constructed, which realizes the time-series source analysis of abnormal behavior and can accurately locate the complete link of "abnormal starting point - propagation chain - final deviation". Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0018] Figure 1 This is a flowchart of the steps of the intelligent assessment method for takeoff and landing stability based on simulator QAR data of the present invention.

[0019] Figure 2 This is an example diagram of reference values ​​for the intelligent assessment method for takeoff and landing stability based on simulator QAR data, as presented in this invention.

[0020] Figure 3 This is an example diagram of the normal distribution of the intelligent assessment method for takeoff and landing stability based on simulator QAR data in this invention. Detailed Implementation

[0021] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] To more clearly explain the intelligent assessment method for takeoff and landing stability based on simulator QAR data of this invention, the following will be combined with... Figures 1 to 3 The steps in the embodiments of the present invention will be described in detail below.

[0024] This invention proposes an intelligent assessment method for takeoff and landing stability based on simulator QAR data. (See [link to relevant documentation]). Figure 1 The method includes the following steps:

[0025] S1. Obtain and decode the historical QAR parameters of all pilots during training on the flight simulator to obtain QAR parameters in CSV format, which are used as the first data. Extract the first data corresponding to different flight states and merge the first data of the same flight segment in different flight stages according to the preset combination identifier field to obtain merged data. The combination identifier field includes the equipment number, the session number, and the training number. The flight stage includes takeoff and landing.

[0026] In this embodiment, QAR stands for QuickAccessRecorder, and the QAR parameters are the parameters stored in the QuickAccessRecorder.

[0027] The first data corresponding to different flight states are extracted, including left pitch control, right pitch control, left roll control, right roll control, rudder, left throttle position, right throttle position, rate of descent, attitude, roll, throttle, instrument airspeed, glide slope deviation, and heading deviation parameters.

[0028] The flight status includes the state of the aircraft being on the ground and the state of the aircraft being in the air. When the aircraft air-to-ground switch signal is on the ground, it is determined that the aircraft is in the state of being on the ground; when the air-to-ground switch signal is in the air and the altitude is greater than 0 feet, the aircraft is in the state of being in the air.

[0029] When the aircraft is on the ground, the first data is extracted from the QAR parameters at time intervals.

[0030] When the aircraft is in the air, the first data is extracted from the QAR parameters at altitude intervals.

[0031] The altitude interval is dynamically adjusted based on the aircraft's radio altitude. The specific adjustment strategy is as follows:

[0032] In the radio altitude range above the first preset threshold, sampling is performed using the first altitude interval;

[0033] In the radio altitude range below the first preset threshold but above the second preset threshold, sampling is performed using a second altitude interval that is smaller than the first altitude interval;

[0034] In the radio altitude range below the second preset threshold but above the third preset threshold, sampling is performed using a third altitude interval that is smaller than the second altitude interval;

[0035] In the radio altitude range below the third preset threshold, a dual-threshold sampling strategy based on time and event is adopted. This strategy is defined as follows: when the time interval between the current sampling point and the previous sampling point reaches a time threshold, or when the absolute value of the altitude difference between the current sampling point and the previous sampling point reaches an event threshold, the current sampling point is considered a valid sampling point.

[0036] For example: when the radio altitude is greater than 200 feet, the altitude interval is 20 feet;

[0037] When 50 feet < radio altitude ≤ 200 feet, the altitude interval is 10 feet;

[0038] When 30 feet < radio height ≤ 50 feet, the height interval is 5 feet;

[0039] At the above altitudes, the data sampling points are all the time points where the altitude error is minimized. The time point where the altitude error is minimized is obtained by the following formula.

[0040] ;

[0041] in, Indicates the current time. The radio altitude is represented by x, which represents the altitude of the current acquisition point. The moment when the minimum absolute value of the difference between the radio altitude and the acquisition point is obtained is the moment when the radio altitude is closest to the acquisition point. Then, the various service parameters at that moment are obtained. The obtained parameters are the various service parameters of the target acquisition point.

[0042] When the radio altitude is ≤30 feet, the pilot will usually pull back on the stick to bring the aircraft into a flare state. At this time, some segments may experience continuous flare at low altitudes. If the previous approach is still used, that is, only the altitude point with the smallest error is selected, it may lead to missed altitude sampling. Therefore, a new approach is adopted in this stage: set an altitude error threshold δ. As long as the radio altitude error is less than the threshold and the time interval between two adjacent sampling points is ≥1 second, the data of that sampling point is extracted. This method allows multiple sampling points at the same altitude (e.g., 15 feet) to avoid missed sampling.

[0043] For example: the time interval Δt from the previous sampling point ≥ 1 second (time threshold) or the height difference |Δh| from the previous sampling point ≥ δ (event threshold);

[0044] The adjacent sampling trigger threshold δ is used to add sampling when there is a significant change in height. Its value is limited to a sampling step size of no more than 5 feet above the (30, 50) feet interval, and is preferably determined as follows:

[0045] ;

[0046] The value of δ is between [1, 3] feet. If you want to sample as much as possible (avoiding missed samples), you can make the value of δ smaller (e.g., δ = 1 ft). If you don't want to sample too much, you can make the value of δ larger (e.g., δ = 3 ft). The above settings ensure that even if the height change is very small within 1 second, a point will be sampled per second through the time threshold. When the height change is large in a short period of time (|Δh|≥δ), the event threshold will sample additional points, thus balancing coverage and responsiveness to rapid changes. This strategy ensures sufficient sampling density in the ≤30 feet range, while satisfying the constraint of δ≤5 feet, and maintaining consistency with the rule of sampling at 5-foot height intervals in the >30 feet range.

[0047] Because the parsed flight segment CSV format QAR parameters are divided into four standard phases for takeoff (GROUND, TKO, INC, CLB) and four standard phases for landing (APP, FNA, LAN, GROUND), the files are scattered, making it difficult to judge the overall situation of a flight. Therefore, the first data of the same flight is merged using the equipment number, flight number, and training number. Among them, the equipment number and flight number of the first data of the same flight are the same, and the training number corresponds to the phase. They are connected and merged in the order of training number.

[0048] For example, training numbers 1→2→3→4 correspond to the takeoff phases GROUND→TKO→INC→CLB in sequence;

[0049] Training numbers 5→6→7→8 correspond to the landing phase APP→FNA→LAN→GROUND in sequence;

[0050] Furthermore, since the training plan does not have all four phases for every flight training, for example, if the aircraft takes off successfully but does not climb a large distance, the INC and CLB phases may be missing. For the missing phases, no observation points are collected, that is, the missing phases in the merged data are marked as null values ​​and will not participate in the overall evaluation.

[0051] S2, based on the merged data, calculate the statistical values ​​of the QAR parameters at different observation points in each flight phase, and determine the reference values ​​of the QAR parameters at each observation point; the statistical values ​​include the mean, standard deviation, and quantile values;

[0052] The formula for calculating the mean is: In the formula, N is the number of observation points, and xi represents the parameter value of the i-th observation point;

[0053] The formula for calculating standard deviation is: In the formula, N is the number of observation points, and xi represents the parameter value of the i-th observation point. This represents the mean;

[0054] Quantile values: First, sort all segment parameter values ​​for the same observation point;

[0055] Xsorted represents the sorted sequence of flight segment parameter values, Xi (i=1,2,⋯,N) represents the i-th flight segment parameter value in the sorted sequence, and N represents the number of elements in the sequence Xsorted;

[0056] Then, the parameter values ​​for specific positions (such as 5%, 10%, 90%, 95%) are obtained;

[0057] like Figure 2 The diagram shows the pilot's manipulation of rudder parameters during the landing phase. It can be observed that pilots do not manipulate the rudder on average during the landing phase, with a mean of 0. However, some pilots will adjust the rudder to align the aircraft with the runway, with a standard deviation of approximately 2.5. Furthermore, the range of values ​​manipulated by most pilots can be used as a reference. After sorting the rudder parameters by magnitude, the specific quantiles for the 5th, 10th, 90th, and 95th percentiles are -5.1, -2.4, 2.4, and 5.1, respectively.

[0058] All parameter reference values ​​for takeoff and landing are stored in a database table. Parameters with the same name are distinguished by the label "type" to indicate whether they are in the takeoff or landing phase. Specifically, type 1 indicates that the parameter information is recorded in the takeoff phase, and type 2 indicates that the parameter information is recorded in the landing phase.

[0059] like Figure 2 As shown: This is the rudder parameter situation during the landing phase. The horizontal axis of the figure is the different radio altitudes ra (e.g., 10 feet, 15 feet) at landing, and the vertical axis is the rudder parameter values; mean, standard deviation, and specific quantiles of 5%, 10%, 90%, and 95%. The green part represents the value of this parameter at the observation point for 80% of the group, i.e., the range of 10%-90%, and the red part represents the value of this parameter at the observation point for 90% of the group, i.e., the range of 5%-95%.

[0060] S3, obtain the QAR parameters of the pilot during training for whom flight stability needs to be determined at different observation points, as the second data; based on the second data, use the improved Lilliefors test method to determine whether the data distribution at the observation points follows a normal distribution; if the normal distribution conforms to the normal distribution shape (see... Figure 3If the normal distribution is true, then the flight stability of the pilot at each observation point is determined based on the normal distribution; otherwise, the flight stability of the pilot at each observation point is determined based on the position of the second data in the overall group distribution interval.

[0061] In this embodiment, based on the second data, an improved Lilliefors test method is used to determine whether the data distribution of the observation points follows a normal distribution. The method is as follows:

[0062] Based on the second data, outliers in the second data are removed using quantile anomaly detection to obtain sample data with outliers removed. Specifically, the second data is first sorted, and then the first quantile Q1, the third quantile Q3, and the interquartile range IQR = Q3 - Q1 are calculated. Based on this, the lower limit for outlier judgment LB = Q1 - 1.5 × IQR and the upper limit UB = Q3 + 1.5 × IQR are determined. If the data value in the second data is less than the lower limit or higher than the upper limit, it is judged as an outlier sample and removed from the sample set.

[0063] Based on the sample data after removing outliers, the Lilliefers test is used to obtain the confidence value p of the data distribution at the observation points. The p-value is used to determine the distribution type of the data at each observation point and to assign a weight to that observation point. Distribution types include those following a normal distribution and those not following a normal distribution; p is the confidence value. For example, if p ≥ 0.1, the data is considered to follow a normal distribution, and the weight of that observation point is set to 1. If 0.01 ≤ p < 0.1, the data is considered to basically follow a normal distribution, and the weight of that observation point is set to 0.5. Both normal and basically normal distributions are determined based on the normal distribution (a reference interval for the normal distribution determined by the mean and standard deviation) to determine the pilot's flight stability at each observation point. If p < 0.01, the data is considered not following a normal distribution. The pilot's flight stability at each observation point is then determined based on the position of the second data point within the overall population distribution interval; that is, the stability is calculated using the quantile method, and the weight of that observation point is set to 1.

[0064] To enhance the influence of key manipulation parameters, the weights of the observation points are further adjusted using parameter sensitivity scoring, the method of which is as follows:

[0065] Based on historical QAR data, with landing performance indicators (such as touchdown overload G value and runway centerline deviation) as prediction targets, and flight parameter sequences as feature vectors, a machine learning model (random forest regression model or gradient boosting tree model, such as XGBoost) is trained, and the feature importance of the model output is extracted as the sensitivity score S of each parameter.

[0066] Using the exponential amplification factor, the observation points of the corresponding parameters are weighted according to the sensitivity score. Nonlinear enhancement is performed, and the enhanced weights are used as the final observation point weights;

[0067] Specifically, the final observation point weights The expression is:

[0068] ;

[0069] Wherein, α>1 is an amplification factor (preferably α=1.2) to ensure that the weight of highly sensitive parameters is exponentially enhanced. This mechanism can automatically increase the influence of key parameters such as pitch angle and descent rate at low altitudes.

[0070] For parameters conforming to a normal distribution, the calculated mean and standard deviation can be used to determine whether the parameter's value range is within the normal range, thus determining the stability of a single observation point. The formula for calculating parameters conforming to a normal distribution is as follows:

[0071] In the formula, This represents the parameter value at the observation point;

[0072] The normal distribution when μ=0 and σ=1 is the standard normal distribution;

[0073] For parameters that are not normally distributed, such as those whose parameter values ​​exhibit a multimodal distribution or have obvious skewness, the stability of the observation point can be evaluated by using the value of a specific quantile as a reference benchmark.

[0074] Whether each parameter conforms to a normal distribution is written to the database after the first successful calculation. Subsequent calculations are not performed; the evaluation criterion is directly selected based on whether the distribution is normal. Specifically, the parameter values ​​of the observed points are compared with the distribution of the overall population. If the value exceeds a preset reference range, the observed point is considered unstable. To quantify this assessment, the data distribution is divided into different intervals, and the following stability is defined:

[0075] If the flight follows or substantially follows a normal distribution, the method for determining the pilot's flight stability at each observation point based on this normal distribution is as follows:

[0076] The intervals are divided based on reference values. The intervals include stable intervals, normal intervals, and unstable intervals, and the stable intervals, normal intervals, and unstable intervals are respectively mapped to stable, normal, or unstable.

[0077] By comparing the parameter values ​​at the observation point with the specified interval, the level of flight stability of the pilot at that observation point is determined to be stable, normal, or unstable.

[0078] If the data does not follow a normal distribution, the method for determining the pilot's flight stability at each observation point based on the position of the second data within the overall population distribution interval is as follows:

[0079] A quantile-based population comparison model is used to obtain the interval of the second data in the overall population distribution. Based on the mapping relationship between the interval and stability, the flight stability of the pilot at each observation point is determined. The quantile-based population comparison model includes the use of the 5%-95% quantile interval.

[0080] The mapping between intervals and stability is as follows:

[0081] The stability interval is defined as the parameter value falling between the 10th and 90th percentiles of the reference value.

[0082] Normal range: The parameter value falls within the 5%-10% quantile or 90%-95% quantile of the reference value;

[0083] Unstable intervals: parameter values ​​< 5th percentile of the reference value or > 95th percentile of the reference value; note that for non-normal distributions, the 5th, 10th, 90th, and 95th percentiles are used, and for normal distributions, the standard deviations corresponding to the 5th, 10th, 90th, and 95th percentiles are used: -5.1, -2.4, 2.4, and 5.1.

[0084] S4. Statistically analyze the flight stability at each observation point. Combine this with preset continuity and proportion thresholds to calculate the pilot's initial stage stability. The method is as follows:

[0085] Count the number of continuously unstable observation points in each stage. If the weight of an observation point is 0.5, count it as 0.5; otherwise, count it as 1.

[0086] Calculate the proportion M of unstable observation points to the total number of observation points in each stage, and calculate the weighted total score of unstable observation points and the weighted total score of all observation points in that stage. The proportion T; if the weight of an unstable observation point is 0.5, then count it as 0.5; otherwise, count it as 1.

[0087] When the ratio M exceeds the preset continuity threshold, the stage is determined to be an unstable state;

[0088] When the ratio T exceeds the preset percentage threshold, the stage is determined to be an unstable state;

[0089] For example: a preset continuity threshold, including 20, and the maximum number of consecutive unstable values ​​does not exceed 20 (excluding 20). The final result of the stability of this parameter stage is stable; otherwise, it is unstable.

[0090] The percentage thresholds include 5%, 20%, and 50%.

[0091] When the ratio T is below 5%, the stage is considered stable.

[0092] When the ratio T is in the range of 5% to 20%, this stage is considered normal;

[0093] When the ratio T is in the range of 20% to 50%, this stage is considered close to normal.

[0094] When the ratio T exceeds 50%, the stage is determined to be unstable; it should be noted that either the ratio T exceeds 50% or the ratio M exceeds 20, either condition is met, and the stage is determined to be unstable.

[0095] S5, when an unstable observation point is identified, a causal network reflecting the dynamic influence relationship between parameters is constructed based on the QAR parameter time series of this flight; based on the causal network, the abnormal propagation path leading to the instability is traced backward, and an abnormality source tracing description is generated. The method is as follows:

[0096] S51, Based on the time series of the QAR parameters, a Vector Autoregression (VAR) model is constructed, wherein the optimal lag order of the model is automatically selected through the information criterion; specifically, for parameters such as pitch angle, rate of descent, airspeed, throttle position, pitch control, and roll control, a VAR(q) model is constructed, where q is the lag order. The Akaike Information Criterion (AIC) is used to calculate the AIC values ​​corresponding to different lag orders q (q ranges from 1 to 10), and the q with the smallest AIC value is selected as the optimal lag order. If there are multiple q corresponding to similar AIC values, combined with the dynamic response characteristics of flight parameters (such as the response delay of control actions is usually 1-3 seconds), q=2 or q=3 is preferentially selected.

[0097] The parameters of each equation in the VAR(q) model are estimated using the ordinary least squares (OLS) method to obtain the coefficient matrix of the lagged terms of each endogenous variable.

[0098] The residual autocorrelation test (LM test) is used to determine whether there is autocorrelation in the model residuals. If the LM test p value is >0.05, it indicates that there is no autocorrelation in the residuals and the model fits well. If autocorrelation exists, the lag order q is adjusted and the model is re-estimated until the model meets the validity requirements.

[0099] S52, based on the constructed vector autoregression model, a Granger causality test is performed to construct a causal directed graph representing the propagation path of influence between parameters. Nodes in the graph represent a flight parameter (such as pitch angle, rate of descent, airspeed, throttle position, etc.), representing the time-series behavior of that parameter within the modeling window. A directed edge from node i to node j indicates that the historical change of parameter i has a statistically significant Granger causal influence on the current change of parameter j. The direction of the edge reflects the order of influence between parameters, and the weight of the edge represents the strength and confidence of the causal relationship. The specific steps are as follows:

[0100] Significance level setting: The significance level of the Granger causality test is set at α=0.05. When the p-value obtained from the test is <0.05, the null hypothesis that "variable X is not a Granger cause of variable Y" is rejected, and it is determined that X "Granger causes" Y.

[0101] Causal strength quantification: The causal strength is calculated based on the F-statistic of the test. The formula is: strength value S=F / (F+10) (where F is the F-statistic of the Granger causality test). The value of S ranges from [0,1]. The closer S is to 1, the stronger the Granger causal influence of X on Y.

[0102] Causal directed graph construction: Using the endogenous variables of the selected VAR model as nodes, if X “Granger causes” Y and S≥0.3 (set a minimum strength threshold to avoid interference from weak causal relationships), then draw a directed edge from X to Y. The thickness of the edge is adjusted proportionally to the strength value S (the larger S is, the thicker the edge), forming a causal directed graph that represents the propagation path of the influence between parameters.

[0103] S53, when an anomaly is detected in a parameter at the observation point, the causal parent node is traced back based on the causal directed graph, and the state of the parent node at historical time is examined to achieve causal back-inference from the final deviation to the starting point of the anomaly, obtain the anomaly propagation path, and generate an anomaly source tracing description; specifically, anomaly parameter identification: real-time monitoring of the QAR parameter time series of the flight to be evaluated, when the value of a parameter Y at time t exceeds the "unstable interval" of the stability judgment, Y is marked as an anomaly parameter at time t;

[0104] Causal parent node backtracking: In a causal directed graph, find the set of all parent nodes {X1, X2, ..., X} that point to Y. n}, that is, all the parameters of Y that "Granger causes";

[0105] Parent node historical state verification: For each parent node Xᵢ, check whether its value at time tk (k is the optimal lag order of the VAR model) exceeds the "unstable interval". If Xᵢ is abnormal at time tk, and the abnormal state is transmitted to Y through causal relationship (i.e., the abnormal time of Xᵢ is earlier than that of Y, and it conforms to the lag response logic of the VAR model), then Xᵢ is marked as a potential abnormal source of Y.

[0106] Anomaly propagation path determination: Repeat the above steps of parent node backtracking and historical state verification, and continue to backtrack the causal parent node of the marked potential anomaly source Xᵢ until the "root node anomaly parameter" that cannot be backtracked further is found, forming a complete anomaly propagation path of "root node anomaly → intermediate propagation node → final anomaly parameter Y".

[0107] Source tracing description generation: Based on the anomaly propagation path, a structured source tracing description is generated, including the anomaly initiation time (root node anomaly occurrence time), root node anomaly parameter name and anomaly degree, intermediate propagation nodes and transmission order, final anomaly parameters and impact, providing accurate basis for pilot training improvement;

[0108] S6, the method for determining the final stage stability risk level by combining the initial stage stability and the anomaly tracing explanation is as follows: based on the anomaly tracing explanation, obtain the type of the root cause of instability; adjust the risk level corresponding to the initial stage stability based on the type of the root cause of instability, and then obtain the final stage stability risk level.

[0109] The types of root causes of stability include critical systemic operational errors, as well as occasional anomalies, general operational errors, and state maintenance deviations.

[0110] When the anomaly tracing reveals that the root cause of instability is a critical systemic operational error, the risk level is increased based on the initial stability level. For example, even if the initial stability is "unstable" (radio altitude > 200ft, sampling interval 20ft; 50ft < altitude ≤ 200ft, interval 10ft; 30ft < altitude ≤ 50ft, interval 5ft; altitude ≤ 30ft, time and event dual-threshold sampling is used, with a time threshold ≥ 1 second or an altitude difference |Δh| ≥ δ (δ ∈ [1, 3]ft); after the Lilliefers test, if the p-value ≥ 0.1, it is considered to follow a normal distribution, and the weight is set to 1; if 0.01 ≤ p-value < 0.1, it is considered to basically follow a normal distribution, and the weight is set to 0.5; if the p-value < 0.01, it is considered not to follow a normal distribution). The data is distributed normally, with a weight of 1. The continuity threshold is 20 consecutive unstable observation points, and the percentage threshold is 50%. That is, if the weighted total score T of unstable observation points exceeds 50%, it is judged as unstable. If any condition is met, it is judged as unstable. If the source tracing explanation reveals that the cause is an occasional, non-critical parameter anomaly, the risk level can be set as "medium". If the source tracing explanation reveals that the cause is a critical systemic control error (such as incorrect pitch control at low altitude), the risk level is set as "high", and targeted training suggestions are provided to the pilots. For example: "If the root node of the source path is 'pitch angle' or 'rate of descent' and there is a continuous anomaly at low altitude (such as <100 feet), it is judged as 'critical systemic control error'. If the root node is 'heading deviation' or the anomaly is brief, it is judged as 'occasional anomaly'."

[0111] S7. Based on the stability level, stage stability, and risk level of a single observation point, perform visualization feedback and database updates, including generating a visualization report containing a heatmap, stage distribution map, and anomaly event axis; and periodically update and check the parameter reference values ​​in the database.

[0112] It should be noted that the Lilliefors test, Vector Autoregression (VAR) model, Granger causality test, and XGBoost algorithm involved are all well-known technologies in this technical field (flight data evaluation, statistical analysis, and machine learning). Their core principles, implementation logic, and application frameworks have been disclosed and therefore not described in detail. The innovative value of this invention is to specifically integrate these well-known algorithms and adapt them to the scenario of "simulator QAR data takeoff and landing stability evaluation": for example, combining the Lilliefors test with "QAR parameter outlier removal" to solve the robustness problem of flight data distribution judgment;

[0113] By combining VAR models, Granger causality tests, and QAR parameter time series, the source of abnormal propagation paths can be traced.

[0114] Combining XGBoost's "feature importance" feature with "flight parameter sensitivity score" to optimize stability assessment weights—these integrations and adaptations address specific problems in flight training assessment scenarios rather than innovating the algorithm itself.

[0115] Although the steps in the above embodiments are described in the above order, those skilled in the art will understand that in order to achieve the effect of this embodiment, different steps do not need to be executed in such an order. They can be executed simultaneously (in parallel) or in a reverse order. These simple variations are all within the protection scope of this invention.

[0116] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0117] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0118] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0119] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for intelligent evaluation of take-off and landing stability based on simulator QAR data, for determining the flight stability of the pilot at each observation point, and then obtaining the stage stability, characterized in that, The method comprises the following steps: S1, obtaining historical QAR parameters of pilots during training on all flight simulators and decoding and analyzing the historical QAR parameters to obtain QAR parameters in CSV format as first data; extracting first data corresponding to different flight states and merging first data of different flight stages of the same flight according to a preset combination identifier field to obtain merged data; S2, based on the merged data, calculating statistical values of QAR parameters of each observation point in each flight stage, and determining reference values of QAR parameters of each observation point; S3, obtaining QAR parameters of a pilot at different observation points during training as second data; based on the second data, using an improved Lilliefors test method to determine whether the data distribution of the observation point conforms to a normal distribution; if yes, determining the flight stability of the pilot at each observation point based on the normal distribution; if no, determining the flight stability of the pilot at each observation point based on the position of the second data in the overall group distribution interval; Based on the second data, an improved Lilliefors test method is used to determine whether the data distribution of the observation point conforms to a normal distribution, and the method is as follows: Based on the second data, the quantile anomaly detection is used to remove the abnormal values in the second data to obtain sample data after removing the abnormal values; Based on the sample data after removing the abnormal values, the Lilliefors test method is used to obtain a confidence value p of the data distribution of the observation point, the distribution type of the observation point data is determined according to the p value, and the weight of the observation point is determined; the distribution type includes conforming to a normal distribution and not conforming to a normal distribution, and the p value is a confidence value; S4, statistically determining the flight stability of each observation point, combining a preset continuity threshold and a proportion threshold, and calculating the preliminary stage stability of the pilot; S5, when an unstable observation point is identified, a causal network reflecting the dynamic influence relationship between parameters is constructed based on the QAR parameter time series of this flight; based on the causal network, the abnormal propagation path leading to instability is traced in reverse, and an abnormal source explanation is generated, and the method is as follows: S51, based on the QAR parameter time series, a vector autoregressive model is constructed, wherein the optimal lag order of the model is automatically selected by information criterion; S52, based on the constructed vector autoregressive model, Granger causality test is performed to construct a causal directed graph representing the influence propagation path between parameters; S53, when an abnormal parameter at an observation point moment is identified, the causal parent node thereof is traced based on the causal directed graph, and the state of the parent node at a historical moment is tested to realize causal backtracking from the final deviation to the abnormal starting point, to obtain an abnormal propagation path and generate an abnormal source explanation; S6, combining the preliminary stage stability and the abnormal source explanation, determining the final stage stability risk level.

2. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 1, characterized in that, The first data corresponding to different flight states is extracted, and the first data of different flight stages of the same flight is merged according to a preset combination identifier field to obtain merged data, comprising: extracting first data corresponding to different flight states; The device number and the game number in the combined identification field corresponding to the first data of the same flight are the same, the training number corresponds to the stage, and the first data is merged in sequence according to the training number; The missing stage in the merged data is marked as null; The flight state includes a state in which the airplane is on the ground and a state in which the airplane is in the air; When the airplane is on the ground, the first data is extracted from the QAR parameters at a set time interval; when the airplane is in the air, the first data is extracted from the QAR parameters at a set height interval; the height interval is dynamically adjusted according to the radio altitude of the airplane; The combined identification field includes a device number, a game number, and a training number; and the flight stage includes takeoff and landing.

3. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 2, characterized in that, The height interval is dynamically adjusted according to the radio altitude of the airplane, and the specific adjustment strategy is: In a radio altitude interval higher than a first preset threshold, a first height interval is used for sampling; In a radio altitude interval lower than the first preset threshold but higher than a second preset threshold, a second height interval smaller than the first height interval is used for sampling; In a radio altitude interval lower than the second preset threshold but higher than a third preset threshold, a third height interval smaller than the second height interval is used for sampling; In a radio altitude interval lower than the third preset threshold, a time and event double threshold sampling strategy is used.

4. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 3, characterized in that, The time and event double threshold sampling strategy is: when the time interval between the current sampling point and the last sampling point reaches a time threshold or the absolute value of the height difference between the current sampling point and the last sampling point reaches an event threshold, the current sampling point is an effective sampling point.

5. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 1, characterized in that, The weight of the observation point further includes adjustment using a parameter sensitivity score, and the method is: Based on historical QAR data, a landing performance index is taken as a prediction target, a flight parameter sequence is taken as a feature vector, a machine learning model is trained, and the feature importance output by the model is extracted as a sensitivity score of each parameter; An exponential amplification coefficient is used to non-linearly enhance the observation point weight of the corresponding parameter according to the sensitivity score, and the enhanced weight is taken as the final observation point weight.

6. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 1, characterized in that, Based on the position of the second data in the overall group distribution interval, the flight stability of the pilot at each observation point is determined, and the method is: A group comparison model based on quantiles is used to obtain the interval of the second data in the overall group distribution, and the flight stability of the pilot at each observation point is determined according to the mapping relationship between the interval and the stability; the group comparison model based on quantiles includes using the 5%-95% quantile interval.

7. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 1, characterized in that, The method for determining the preliminary stage stability of the pilot is: The number of consecutive unstable observation points in each stage is counted; The proportion M of the number of unstable observation points in each stage to the total number of observation points in the stage is calculated, and the proportion T of the weighted total score of the unstable observation points to the weighted total score of all observation points in the stage is calculated; When the proportion M exceeds a preset continuity threshold, the stage is determined to be unstable; When the proportion T exceeds a preset proportion threshold, the stage is determined to be unstable.

8. The method for intelligent evaluation of take-off and landing stability based on analog machine QAR data according to claim 1, characterized in that, The method for determining the final stage stability risk level by combining the preliminary stage stability with the abnormality traceability explanation is as follows: acquiring the type of the unstable root cause according to the abnormality traceability explanation; and adjusting the risk level corresponding to the preliminary stage stability according to the type of the unstable root cause, thereby obtaining the final stage stability risk level. The type of the stable root cause includes a critical systematic manipulation failure.

Images