A flight potential prediction method based on expert experience and dynamic image generation technology

Abstract

This invention provides a flight potential prediction method based on expert experience and dynamic profile generation technology, belonging to the field of flight potential prediction technology. The invention first constructs a multi-dimensional evaluation table including flight capability dimensions and non-capability factor dimensions, with two experts independently scoring to form an initial subjective assessment; it then establishes a mapping relationship between expert assessment scores and key objective flight parameters, parametrically calibrating the initial subjective assessment; it aggregates the calibrated expert assessments, objective parameter aggregation indicators, and psychological test data, forming a flight potential classification model based on potential profile analysis; finally, it generates and dynamically updates the flight potential profile, constructs a future flight performance prediction model, and dynamically corrects it. This method improves the stability, completeness, and predictive accuracy of flight potential assessment.

Description

Technical Field

[0001] This invention relates to the field of flight potential prediction technology, and in particular to a flight potential prediction method based on expert experience and dynamic profile generation technology. Background Technology

[0002] In the selection and training evaluation of flight personnel, a combination of expert scoring and simulated flight assessment is typically used to evaluate candidates. Currently, flight experts score candidates based on their operational performance, reaction time, and mission completion during mission execution. The simulated flight system records flight parameters such as attitude, flight path deviation, and operational timing. In some scenarios, psychological assessment results are also incorporated for supplementary judgment. This method can meet the basic assessment and phased screening needs in flight training.

[0003] As the data acquisition capabilities of flight simulation training equipment continue to improve, training evaluation work places higher demands on the detail and continuity of assessment results. In practical applications, in addition to focusing on the score of a single assessment, it is also necessary to analyze the performance changes of the evaluated subject across multiple training phases to understand their capability development trends. Furthermore, it is essential to incorporate objective flight parameters for comparison based on expert scoring to reduce the impact of scoring discrepancies on the evaluation results, thereby improving the stability and usability of the training evaluation outcomes.

[0004] However, existing technologies still have the following problems: First, expert scores are easily influenced by personal experience and scoring habits, and the scoring results may differ between different experts; second, there is a lack of systematic correspondence between expert scores and objective flight parameters, making it difficult to correct subjective scores; third, existing evaluation results are mostly single or interim conclusions, making it difficult to continuously update them based on subsequent training data; and fourth, there is insufficient comprehensive utilization of flight capability factors and non-capability factors such as psychology, making it difficult to form potential evaluation results that adjust with data changes during training and further use them for subsequent flight performance prediction. Based on this, it is necessary to propose a dynamic prediction method for flight potential that can combine expert scores, objective flight parameters, and subsequent training data. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a flight potential prediction method based on expert experience and dynamic profile generation technology. By using independent scoring by two experts, calibration of objective flight parameters, multi-dimensional evaluation data modeling, and dynamic profile updates, the stability, completeness, and subsequent predictive value of flight potential assessment results can be improved.

[0006] To achieve the above objectives, the present invention provides the following solution: A flight potential prediction method based on expert experience and dynamic profiling generation technology includes: Construct a multi-dimensional evaluation table that includes flight capability dimensions and non-capability factor dimensions; Based on the multidimensional evaluation form, two experts independently score the performance of the same student in the same subject, forming two original subjective evaluations; Establish a mapping relationship between expert evaluation scores and key objective flight parameters, and perform parameterized calibration on two original subjective evaluations according to preset rules to obtain calibrated expert evaluations; After calibration, expert evaluations, aggregated objective parameters, and psychological test data are collected to form multi-dimensional evaluation data. Based on the multi-dimensional evaluation data, potential profile analysis is conducted to form a flight potential classification model. The flight potential classification model is used to classify the flight potential of trainees and generate flight potential profiles. The phased flight performance and new assessment data are input into the flight potential classification model to dynamically update the flight potential profiles and obtain phased update results. Based on flight potential profiles and phased update results, a predictive model for future flight performance is constructed. The predictive model is dynamically corrected based on the newly added flight performance data in each round. Based on the dynamically corrected predictive model, the trainee's performance level, risk propensity, and adaptability in future specific flight missions or training phases are predicted.

[0007] The present invention discloses the following beneficial effects: (1) This invention constructs a multi-dimensional evaluation table that includes flight ability dimensions and non-ability factor dimensions, and has two experts independently score the same trainee under the same subject. This transforms the originally scattered expert experience judgments into evaluation results that can be recorded separately, thereby improving the standardization and comparability of the evaluation process from the source. Compared with the existing practice of relying solely on the experience judgment of a single expert, this invention helps to reduce the deviation caused by the fluctuation of scores due to personal habits and improves the problem of insufficient consistency in expert scores.

[0008] (2) By establishing a mapping relationship between expert evaluation scores and key objective flight parameters, and by parametrically calibrating the two original subjective evaluations, this invention can correlate subjective scores with objective performance during flight, thereby making the calibrated expert evaluations more closely reflect actual operational performance. Compared to the prior art where subjective scores and objective flight parameters are separate and difficult to mutually correct, this invention can improve the reliability and stability of the evaluation results.

[0009] (3) This invention gathers multi-dimensional assessment data by combining expert evaluations after calibration, aggregated objective parameter indicators, and psychological test data. Based on this multi-dimensional assessment data, a potential profile analysis is conducted to form a flight potential classification model, which can incorporate flight ability factors and non-ability factors into the same evaluation process. Compared with the existing technology, which does not make sufficient comprehensive use of ability factors and non-ability factors such as psychology, this invention can more completely reflect the potential characteristics of trainees and reduce the limitations of making judgments based on a single score or a single-dimensional indicator.

[0010] (4) This invention generates a flight potential profile using a flight potential classification model, and inputs the phased flight performance and new evaluation data into the flight potential classification model to dynamically update the flight potential profile, enabling the evaluation results to be continuously adjusted as the training progresses. Compared with the existing technology where the evaluation results are mostly single or phased static conclusions, this invention can reflect the performance changes of trainees at different training stages, improving the continuous use value of flight potential evaluation results in training management.

[0011] (5) This invention constructs a predictive model for future flight performance based on flight potential profiles and phased update results, and dynamically corrects the predictive model according to the newly added flight performance data in each round. This can continuously improve the matching degree between the prediction results and the actual performance as subsequent data continues to increase. Compared with the problem of existing technologies that it is difficult to further use the previous assessment results for predicting the performance of future flight missions or training phases, this invention can provide a continuously updated data foundation for predicting performance level, risk propensity, and adaptability. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the classification of student potential prototypes provided in an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] The purpose of this invention is to provide a flight potential prediction method based on expert experience and dynamic profile generation technology. By parametrically calibrating expert scores and combining them with multi-dimensional data to form a dynamically updatable flight potential profile, the stability, information completeness, and predictive ability of flight potential evaluation results can be improved.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a flight potential prediction method based on expert experience and dynamic profile generation technology, comprising: Step 100: Construct a multi-dimensional evaluation table that includes flight capability dimensions and non-capability factor dimensions; Step 200: Based on the multidimensional evaluation table, two experts independently score the performance of the same student in the same subject, forming two original subjective evaluations; Step 300: Establish the mapping relationship between expert evaluation scores and key objective flight parameters, and perform parameterized calibration on the two original subjective evaluations according to preset rules to obtain the calibrated expert evaluations; Step 400: Gather expert evaluations, aggregated objective parameters, and psychological test data after calibration to form multi-dimensional evaluation data, and conduct potential profile analysis based on the multi-dimensional evaluation data to form a flight potential classification model; Step 500: Use the flight potential classification model to classify the students' flight potential and generate flight potential profiles. Input the phased flight performance and new assessment data into the flight potential classification model to dynamically update the flight potential profiles and obtain phased update results. Step 600: Based on the flight potential profile and the results of phased updates, construct a predictive model for future flight performance. Dynamically revise the predictive model according to the new flight performance data in each round, and predict the trainee's performance level, risk propensity and adaptability in future specific flight missions or training phases based on the dynamically revised predictive model.

[0018] Specifically, in this embodiment, step 100 is used to complete the structured collection of flight expert experience, specifically by constructing a multi-dimensional evaluation table that includes flight capability dimensions and non-capability factor dimensions. The multi-dimensional evaluation table refers to an evaluation carrier that breaks down, records, and configures observation content that originally relied on expert experience judgment according to predetermined dimensions. Its function is to provide a unified evaluation basis and recording format for two experts to independently score the same student in the same subject, avoiding deviations in scoring standards due to inconsistent focus among different experts. The expert observation clues refer to behavioral, operational, or emotional behaviors corresponding to each evaluation item that can be directly observed, judged, and recorded during simulated flight or interviews. These clues constrain the scope of expert scoring and improve scoring consistency. Preferably, the multi-dimensional evaluation table includes at least the evaluation item settings, corresponding expert observation clues, and scoring and annotation recording locations. "Annotation" includes textual recording of observed key behaviors, abnormal behaviors, deviation correction processes, or situational trigger points. "Retrospective scoring and annotation" refers to experts reviewing and completing scoring and annotation based on the recorded information from the simulated flight process after the process ends. Its function is to supplement key behaviors that may have been missed during real-time observation.

[0019] In this embodiment, when setting evaluation items for the flight capability dimension and evaluation items for the non-capability factor dimension, the evaluation content is organized hierarchically according to the flight training potential assessment scenario. Preferably, evaluation items for comprehension and acceptance, memory ability, control ability, attention ability, spatial orientation, adaptability, and situational awareness are set in the flight capability dimension; evaluation items for emotional control, willpower, motivation, and personality are set in the non-capability factor dimension, as shown in Table 1.

[0020] Table 1. Evaluation Items of the Dimensional Evaluation Table Furthermore, corresponding expert observation cues are configured for each evaluation item, enabling experts to make judgments based on specific performance rather than generalizing evaluations. For example, for the control ability evaluation item, expert observation cues could correspond to the timing, amplitude, and directional accuracy of throttle, stick, and pedal control, as well as the coordination of hands and feet; for the attention ability evaluation item, expert observation cues could correspond to the detection and correction of deviations in heading, speed, and altitude during ascent, descent, and turning practice, as well as the balance between aircraft status, data, and control actions; for non-ability factors such as emotional control, willpower, motivation, and personality, expert observation cues can be obtained through observation throughout the simulated flight process and interviews, used to record trainees' emotional and behavioral reactions, persistence, career interests, and attribution styles when the difficulty increases, mistakes occur, or they are questioned or pursued. Through the above evaluation item settings and observation cue configuration, the multi-dimensional evaluation table can unify ability factors and non-ability factors into a single evaluation entry point, providing a structured input basis for the subsequent steps to form two original subjective assessments.

[0021] In this embodiment, the evaluation items are configured with corresponding expert observation cues to form the multi-dimensional evaluation table, which is then provided to two experts for independent scoring of the same student and the same subject. Specifically, experts can perform real-time scoring and annotation during simulated flight, or retrospective scoring and annotation based on the same flight data and performance records after the simulated flight. Regardless of whether a real-time or retrospective approach is used, the same evaluation items and expert observation cues from the same multi-dimensional evaluation table are used to ensure consistency in the evaluation object, evaluation dimension, and observation basis between the two original subjective assessments. Preferably, the output of the multi-dimensional evaluation table after its formation serves as the direct input basis for subsequent step 200, enabling step 200 to complete independent scoring by two experts and form two original subjective assessments based on the table. Simultaneously, since the multi-dimensional evaluation table clearly defines the evaluation items and observation cues for both ability and non-ability factors, the parametric calibration in subsequent step 300 and the multi-dimensional evaluation data aggregation in step 400 can be processed under a unified dimensional caliber.

[0022] Further, in this embodiment, step 200 is used to generate two original subjective assessments and complete a two-expert consistency check. Specifically, this embodiment calls the multi-dimensional evaluation table generated in step 100. Based on the same flight data and performance records, two experts independently score the performance of the same trainee in the same subject, generating two original subjective assessments. The same flight data and performance records refer to a set of records from the same simulated flight mission, the same trainee, the same subject, and the same time period. Their function is to ensure that the observation basis of the two experts is consistent when they score. Among them, the flight data may include records of heading, speed, altitude, attitude changes, control input timing, and deviation correction process, etc., and the performance records may include real-time annotations or retrospective annotations generated according to the expert observation clues in step 100. The original subjective assessment refers to the scoring results and accompanying annotations given by the experts directly based on the multi-dimensional evaluation table before the calibration of key objective flight parameters. Its function is to serve as input for subsequent consistency analysis and parameterized calibration. In this embodiment, a unified scoring system is preferably used for recording. For example, the scores of each evaluation item are uniformly converted into a percentage system before entering the subsequent processing, so that the scores of different dimensions can be tested and calibrated on the same numerical scale.

[0023] In this embodiment, the correlation analysis, difference test, consistency index calculation, consistency warning, and cross-checking in step 200 are performed sequentially to obtain stable input that can be used in step 300. Specifically, this embodiment first performs correlation analysis and difference test on the two original subjective assessments and calculates the consistency index for each assessment dimension. The consistency index refers to a quantitative indicator used to characterize the degree of consistency between the scores of the two experts. Its function is to provide a basis for judging whether to trigger a consistency warning, and may include one or more of the following: intragroup correlation coefficient, score deviation value, and key item deviation ratio. Preferably, an intragroup correlation coefficient of 0.75 or above is considered basically consistent, and an intragroup correlation coefficient of 0.85 or above is considered highly consistent. The score deviation value can be calculated as the absolute value of the percentage score difference, preferably set as a graded judgment range of 3 to 15 points. For example, a difference of no more than 5 points is considered small, and a difference of more than 10 points but no more than 15 points is considered to require key review. In the case where the difference in key item scores exceeds a preset threshold, this embodiment triggers a consistency warning. The key items refer to evaluation items that significantly impact subsequent flight potential assessments, such as controllability, responsiveness, and situational awareness. The consistency warning refers to the review prompt information generated in this embodiment, which notifies experts to review discrepancies rather than directly overwriting the original score. Preferably, the preset threshold for key items can be configured separately for each subject, for example, set to 8 to 12 points on a percentage scale, specifically 10 points. After triggering, two experts perform cross-review based on the original scoring criteria and performance records. The cross-review refers to the two experts comparing and explaining the scoring criteria for discrepancies and reconfirming the scores or annotations. Its function is to reduce accidental biases caused by differences in observation angles. The record after cross-review serves as one of the inputs for step 300.

[0024] In this embodiment, step 300 first establishes a mapping relationship between expert evaluation scores and key objective flight parameters, and then determines whether the two original subjective evaluations are consistent with the operational performance reflected by the key objective flight parameters. The key objective flight parameters refer to a set of flight parameters that can directly or indirectly characterize the performance level of the corresponding evaluation item, and their function is to provide an objective reference for subjective scoring. For example, parameters such as heading deviation, speed deviation, altitude deviation, smoothness of control input, and number of correction overshoots can be used for control capability; parameters such as reaction time after an abnormal state occurs, delay of the first corrective action, and number of consecutive corrections can be used for responsiveness; parameters such as correct action response after a prompt event is triggered and the accuracy of state recognition can be used for situational awareness. The mapping relationship refers to a correspondence table, weight configuration, or rule set between the evaluation item and its corresponding key objective flight parameters, and its function is to establish a comparable relationship between the expert evaluation scores of different evaluation items and the operational performance represented by the corresponding parameters. In this embodiment, it is preferable to first process the key objective flight parameters to be of the same dimension and directional, and then form an objective reference score according to the evaluation item for comparison and judgment with the two original subjective evaluations. To ensure that the determination of serious discrepancies, high consistency, and significant deviations has a parameter basis, this embodiment preferably sets the following ranges: When the difference between the expert score and the objective reference score does not exceed 8 points, and the directional judgment of the corresponding parameter for the evaluation item is consistent, it is determined to be of high consistency; when the difference reaches 20 to 30 points, or when two or more key objective flight parameters corresponding to the same evaluation item simultaneously fall into the abnormal range, it is determined to be of serious discrepancy; when the difference exceeds 12 points, or when the difference still exceeds 10 points after two consecutive reviews, it is determined to be of significant deviation; when both original subjective assessments are higher or lower than the objective reference score relative to it, and their respective differences reach 8 points or more, it is determined to be of consistent deviation direction; when both original subjective assessments and objective reference scores meet the high consistency requirement, but the difference between the scores of the two experts is in the range of 3 to 8 points, it is determined to be a reasonable difference. The above ranges can be pre-configured by subject or training phase and recorded in the parameter rules of this embodiment.

[0025] In a preferred embodiment, when step 300 determines that there is a significant deviation, a rule-driven parametric calibration method is used to parametrically calibrate the two original subjective assessments to obtain a calibrated expert assessment. The parametric calibration refers to a process of quantitatively adjusting or synthesizing the score value of the original subjective assessment based on preset rules and combined with key objective flight parameters. Its function is to make the output score more consistent with objective operational performance. Specifically, when one original subjective assessment is seriously inconsistent with the key objective flight parameters, and the other original subjective assessment has a high degree of consistency with the key objective flight parameters, this embodiment prioritizes adopting the original subjective assessment with higher consistency and performs weighted adjustments on the abnormal scores to obtain the calibrated expert assessment. Preferably, the weight of the original subjective assessment with high consistency can be set to 60% to 85%, the weight of the objective reference score formed based on the key objective flight parameters can be set to 15% to 40%, and the weight of the seriously inconsistent original subjective assessment can be reduced to 0% to 20%. Specific examples include 70%, 25%, and 5%. When both original subjective assessments deviate from the key objective flight parameters in the same direction, this embodiment performs synchronous calibration on the two original subjective assessments according to parameter rules. Preferably, the two scores are adjusted by 3 to 12 points respectively in the same calibration direction, or scaled according to the same proportional range, to obtain the calibrated expert assessment. When both original subjective assessments are highly consistent with the key objective flight parameters but have reasonable differences, this embodiment retains the reasonable differences and generates an interval score or a composite score based on the reasonable differences as the calibrated expert assessment. The interval score refers to the score range formed by the lower and higher scores, and its function is to retain the differences in expert focus; the composite score refers to the single score formed by weighted synthesis of the two scores, and its function is to be directly called by subsequent steps 400. Preferably, the interval width is controlled within the range of 3 to 8 points; if it exceeds this range, a re-review or re-determination of significant deviations is triggered.

[0026] In another preferred embodiment, the parametric calibration in step 300 is implemented using a regression model trained on historical data to automatically calibrate the new assessment and output the calibrated expert assessment. The regression model is a numerical prediction model used to predict calibration scores based on input features; its function is to learn the deviation patterns between the original subjective assessment and key objective flight parameters. This embodiment discloses the regression model using method steps, specifically including training data construction steps, model training steps, and model application steps. In the training data construction step, historical samples are collected and training samples are constructed separately for each evaluation item. The input of each training sample includes the set of objective parameters corresponding to the same evaluation item and the original expert score for that evaluation item. The output includes the standard score for that evaluation item after arbitration by two authoritative experts, thereby ensuring the correspondence between the input and output in terms of the evaluation item's scope. In the model training step, the regression model is trained and validated using a training set, validation set, and test set partitioning method, with preferred partition ratios of 70%, 15%, and 15%. Exemplarily, the number of training epochs can be set to 50 to 300, the learning rate can be set to 0.0001 to 0.01, the batch size can be set to 16 to 256, and the early stopping condition can be set to the validation set error not improving within 5 to 20 consecutive epochs. In the model application step, when a new evaluation arrives, the objective parameter set corresponding to the target evaluation item and the original expert scores are input into the trained regression model according to the same input structure as in the training phase, obtaining the calibrated expert evaluation for that evaluation item. This calibrated expert evaluation is then provided to step 400 for multi-dimensional evaluation data aggregation. By maintaining consistency in the mapping of evaluation items between the training and application phases, this embodiment ensures a clear intrinsic relationship between the input and output of the regression model, avoiding calibration distortion caused by mixing evaluation items.

[0027] In this embodiment, to quantify the consistency between the original expert scores and the operational performance reflected by key objective flight parameters, an evaluation item-level deviation index can be defined as the basis for calculating high consistency, serious discrepancy, and significant deviation in step 300. Specifically, it can be expressed as follows: in, For the first Experts at the The deviation index on each evaluation item is used to characterize the overall degree of deviation between the expert's original subjective assessment and the objective reference score; For the first Experts on the first The original subjective evaluation score given by each evaluation item; For the first The objective reference score corresponding to each evaluation item is generated by processing the key objective flight parameters corresponding to the evaluation item with the same dimensions and the same direction. For the first The deviation judgment threshold for each evaluation item is used to normalize the scoring deviation of different evaluation items. It is preferred to pre-configure a value between 8 and 12 points for each subject, and 10 can be used as an example. For the first Experts at the The number of abnormal parameters on each evaluation item that are inconsistent with the judgment results of key objective flight parameters; For the first The total number of key objective flight parameters corresponding to each evaluation item is preferably an integer between 2 and 6, with 4 being an example; when Not greater than 1 and simultaneously satisfying A score of 8 or less can be considered a high degree of consistency. Reaching 2 or above When the deviation index reaches the range of 20 to 30 points, it can be judged as a serious discrepancy. When the deviation index continues to be higher than the preset warning value and does not decrease to the preset range after review, it can be judged as a significant deviation.

[0028] In this embodiment, to generate calibrated expert evaluations under a rule-driven path, the original subjective evaluation scores of two experts and the objective reference scores can be adaptively fused for consistency. Specifically, the following formula can be used: in, For the first The average deviation index of two experts corresponding to each evaluation item is used to characterize the overall deviation level of that evaluation item in the current sample. For the first The calibrated expert evaluation scores of each evaluation item are used as the output of step 300 and are collected in step 400 to form multi-dimensional evaluation data. To prevent the stability constant from having a denominator of 0, it is preferable to take one value between 0.1 and 1, for example, 0.5.

[0029] In this embodiment, step 400 first aligns and aggregates the calibrated expert evaluation, objective parameter aggregation indicators, and psychological assessment data output from step 300 according to the same trainee, the same subject, and the same assessment round, forming multi-dimensional evaluation data, which serves as the input sample for potential profile analysis. Indicators with different dimensions and opposite directions are first standardized in both direction and value to ensure that the flight capability dimension and the non-capability factor dimension are comparable within the same statistical modeling space. Preferably, the standardized multi-dimensional evaluation data matrix elements are constructed using the following formula: in, For the first The student in the first Standardized values ​​on each evaluation dimension are used as modeling inputs for potential profile analysis; For the first The student in the first The original aggregated values ​​for each assessment dimension are derived from the corresponding dimension data in calibrated expert assessments, aggregated objective parameter indicators, and psychological test data. For the first The mean of each evaluation dimension in the current modeling sample set; For the first The standard deviation of each evaluation dimension in the current modeling sample set; The meaning follows the previous definition. To prevent the stability constant from having a denominator of 0, it is preferred to take one value between 0.01 and 1 in this step. For example, 0.1 is taken. For student index, the value ranges from 1 to... ; For evaluating the dimension index, the value ranges from 1 to... ,in It should at least cover all evaluation items or aggregated indicators involved in the modeling under the dimensions of flight capability and non-capability factors, so that there is a clear dimensional correspondence between the input data and the subsequent trait category output.

[0030] In this embodiment, potential profile analysis is applied to perform statistical modeling based on the aforementioned multi-dimensional evaluation data to identify potential trait categories. The basic parameters of the flight potential classification model are formed through category number search and model optimization. Preferably, Gaussian mixture modeling is used for potential profile analysis, and the number of categories is selected in conjunction with information criteria. Specifically, the following formula can be used: in, The number of categories is The sample likelihood function value at that time is used to characterize the degree of fit of the current potential profile model to the multi-dimensional evaluation data; The number of categories is The information criterion value at that time is used to make a balance between the goodness of fit and the complexity of the model; For the number of potential trait categories, it is preferable to search one by one within the range of 2 to 8. For the first The prior proportion parameter of each trait category satisfies the condition that it is non-negative and the sum of the prior proportions of each category is 1; For the first The trait category in the first Mean parameters across each evaluation dimension; For the first The trait category in the first Variance parameters on each evaluation dimension; The number of categories is The number of free parameters in the time model; in this embodiment, the expected maximization iterative estimation is preferred to estimate the above parameters, and the convergence stopping condition is set to the log-likelihood increment of two adjacent iterations not exceeding to One value in the example, taking one of the values. Meanwhile, the maximum number of iterations is set to one of the values ​​between 100 and 500, with 300 iterations used in the example, to ensure that the model training process is reproducible and the parameter estimation is sufficient.

[0031] In this embodiment, after determining the number of categories and completing parameter estimation, the trainee is identified by trait category based on the posterior attribution probability, and prototype classification is completed from the dimensions of flight ability and non-ability factors. The posterior attribution probability and prototype attribution can be expressed by the following formula: in, For the first The student belongs to the first The posterior attribution probability of each trait category is used to characterize the degree of matching between the trainee and each potential prototype; For the first The prototype segmentation results for each student; The index is used for category summation; the meanings of the remaining parameters follow the aforementioned definitions; in this embodiment, the maximum posterior assignment probability is preferably used as the prototype partition confidence level, and the prototype partition confidence threshold is set to one value between 0.60 and 0.85, for example, 0.70. When the maximum posterior assignment probability of a trainee is lower than the threshold, the trainee is marked as a low-confidence sample and enters manual review or the next round of sample expansion training, so as to reduce the impact of boundary samples on the stability of prototype definition, thereby ensuring that the output of potential profile analysis is consistent with the classification rules in the subsequent flight potential classification model.

[0032] In this embodiment, after obtaining the prototype segmentation results, descriptive labels and core feature vectors are further generated for each prototype. A flight potential classification model is then formed based on the prototype, descriptive labels, and core feature vectors. The core feature vectors are preferably generated in a way that balances inter-class differences and intra-class stability, and can be specifically generated using the following formula: in, For the first The prototype in the first The core feature vector components on each evaluation dimension are used to characterize the distinguishing strength of the prototype on the corresponding dimension and serve as prototype feature parameters for subsequent flight potential classification models. For the first The mean of each evaluation dimension in all modeling samples; For the first The stabilization variance constant for each evaluation dimension is preferably one value between 0.001 and 0.1, with an example of 0.01; To be classified as the The number of samples for each prototype; in this embodiment, the core feature vector components corresponding to each prototype are sorted from largest to smallest, and the top 3 to 8 evaluation dimensions are selected as the core discriminant dimensions of the prototype. Descriptive labels are generated by combining the performance of the core discriminant dimensions in the flight capability dimension and the non-capability factor dimension. Finally, a flight potential classification model is formed, which includes the number of categories, the prior proportion of each category, the dimension parameters of each category, the core feature vector, and the label mapping rules. This allows the model parameters output in step 400 to be directly called in step 500, and makes the input multi-dimensional evaluation data and the output prototype category, label, and core feature vector have a clear and implementable correspondence.

[0033] In this exemplary embodiment, the objective parameter aggregation index refers to a dimensional numerical value formed by statistically summarizing key objective flight parameters for the same trainee, the same subject, and the same assessment round according to a preset time window or task stage. Its function is to convert the original flight process parameters into input items that can be used for modeling together with calibrated expert assessment and psychological test data. Psychological test data refers to the quantitative results obtained through pre-selected psychological assessment tools that match the flight training scenario. It is preferably converted into sub-item scores or subscale scores corresponding to the non-ability factor dimensions. Its function is to characterize the performance of non-ability factors such as emotional control, willpower, motivation, and personality. Potential profile analysis in this embodiment refers to a statistical modeling method for classifying continuous multi-dimensional assessment data. Its output trait categories refer to potential groups with similar performance patterns in both the flight ability and non-ability factor dimensions. The prototype refers to a trainee type representation formed based on the trait categories, used for subsequent flight potential classification and mapping. The process involves several steps: Descriptive labels are type descriptions generated based on the prototype's performance across core discriminative dimensions. Their function is to improve the interpretability of classification results and the readability of subsequent applications. Core feature vectors are vectorized parameters representing the distinguishing strength of each prototype across various evaluation dimensions. Their function is to serve as prototype parameters for the flight potential classification model, which can then be used in subsequent steps. Low-confidence samples are samples whose maximum posterior probability of attribution does not reach a pre-set confidence threshold. Their function is to identify boundary samples and avoid directly using them for stable prototype definitions. Sample augmentation training refers to the process of supplementing new trainee samples or low-confidence samples in subsequent rounds and then re-executing the latent profile analysis parameter estimation. Its function is to improve the stability of prototype segmentation. Label mapping rules are a set of rules that map prototype parameters and core discriminative dimension performance to descriptive labels. Their function is to ensure a consistent correspondence between the prototype category, descriptive labels, and core feature vectors output in step 400, and to allow them to be directly used in step 500.

[0034] Optionally, in this embodiment, step 400 corresponds to the basic modeling process for generating a potential profile based on a latent variable model. Specifically, this embodiment collects a large amount of multi-dimensional assessment data from trainees. This multi-dimensional assessment data includes at least calibrated expert assessments, aggregated objective parameter indicators, and psychological test data. This multi-dimensional assessment data is used as input samples for potential profile analysis. Specifically, the calibrated expert assessments characterize the scoring results of flight capability dimensions and non-capability factor dimensions after parameter calibration in step 300. The aggregated objective parameter indicators characterize the statistical results of key objective flight parameters within a preset time window or mission phase. The psychological test data characterizes the quantitative results of non-capability factors such as emotional control, willpower, motivation, and personality. This embodiment applies statistical modeling methods such as potential profile analysis to jointly model surface-level, related observed variables to identify a few potential and independent trait categories hidden behind these observed variables. The identified trait categories serve as the basis for subsequent prototype segmentation, descriptive label generation, and core feature vector construction, thereby forming the category structure parameters of the flight potential classification model.

[0035] In this embodiment, after identifying potential trait categories, step 400 further categorizes trainees into prototypes based on both flight ability and non-ability factors, and generates descriptive labels and core feature vectors for each prototype. Specifically, this embodiment categorizes trainees into four combined prototypes: high ability and strong non-ability factors, high ability and weak non-ability factors, low ability and weak non-ability factors, and low ability and strong non-ability factors. Each combined prototype is further subdivided into descriptive labels with different typical characteristics. The high ability and strong non-ability factor prototypes may include the precise and stable type and the composed type, representing trainees with both excellent talent and temperament, who are the focus of training. The high ability and weak non-ability factor prototypes may include the arrogant and complacent type and the psychologically volatile type, representing trainees with outstanding talent but psychological or motivational shortcomings, who are prone to becoming problem trainees, and who are the focus of attention. Prototypes with weak performance factors can include the bewildered and anxious types, characterized by significant shortcomings in spatial orientation, situational awareness, and control abilities. They often exhibit busyness, disorganization, forgetfulness, severely irrational attention allocation, vague motivation, lack of confidence, and may be introverted or overly dependent. These individuals are prone to giving up easily when faced with difficulties and require focused assessment of their flight potential. If no improvement is seen, early diversion should be considered. Prototypes with low performance but strong non-performance factors can include the diligent and hardworking types, characterized by average talent but excellent psychological qualities and a diligent attitude, representing potential late bloomers. Their reactions and controls may be slow, but once mastered, they are relatively solid and rarely make repetitive errors. This embodiment generates descriptive labels and core feature vectors for each type based on the above prototype classification results, serving as the direct parameter basis for generating and updating the flight potential profile in step 500. Figure 2 As shown, Figure 2The invention illustrates the potential prototype category information formed by combining and classifying trainees based on flight ability dimensions and non-ability factors. These categories include the precise and stable type, the composed type, the arrogant and complacent type, the psychologically fluctuating type, the bewildered type, the tense and out-of-control type, the diligent and capable type, and the slow-burning and steady type. This information is used to illustrate the prototype classification results when generating trainee flight potential profiles based on potential trait category identification in this invention.

[0036] Further, in this embodiment, step 500 is used to classify the student's flight potential, generate a flight potential profile, and perform phased dynamic updates based on the flight potential classification model formed in step 400. Specifically, this embodiment first calls the prototype category parameters, descriptive labels, and core feature vectors output in step 400 to classify the current student's flight potential and obtain classification results; then, a flight potential profile is generated based on the classification results. The flight potential profile refers to a structured representation result for a single student, including at least the potential profile type, the feature strength corresponding to each evaluation dimension, the classification confidence level, and the corresponding evaluation round identifier. Its function is to provide intermediate results for subsequent phased tracking updates and future performance prediction in step 600. The phased flight performance refers to the phased assessment score or phased task completion level quantification result generated by the same student in subsequent training phases. The newly added evaluation data refers to data newly collected in the subsequent training phases that can be mapped to the flight ability dimension and non-ability factor dimensions, preferably including one or more of the following: newly calibrated expert evaluation, newly added objective parameter aggregation indicators, and newly added psychological assessment data. In this embodiment, the phased flight results and the newly added evaluation data are aligned according to the same student and the same phase round before being input into the flight potential classification model, so as to keep the dimensionality of the input in step 500 consistent with that of the prototype parameters in step 400, and output the phased update results for step 600 to call.

[0037] In this embodiment, to ensure that the impact of newly added stage information on classification probability has a calculation basis, this embodiment uses the prototype parameters and historical attribution probabilities obtained in step 400 to evaluate the students' performance in the first stage. The classification probabilities during round updates are incrementally updated. The preferred method for constructing stage evidence items and calculating updated classification probabilities is as follows: in, For the first The student in the first During the update round, relative to the first round Each prototype category has a phased evidence item, which is used to comprehensively characterize the degree of support of historical attribution status, phased flight performance, and newly added assessment data for the current category attribution; For the first The student in the first The updated class assignment probability is obtained by taking the posterior assignment probability output in step 400 in the initial round. For the first The student in the first The probability of category assignment after each round of updates; For the first The student in the first The phased flight performance of the wheel; For the first The benchmark value of the phased flight performance corresponding to each prototype category is obtained by statistical analysis of the phased flight performance of the historical high-confidence sample of that prototype. For the first The dispersion parameter of the stage flight performance corresponding to each prototype category is obtained by statistical analysis of the variance of the stage flight performance of the historical high-confidence samples. For the first The student in the first The newly added evaluation data in the first round is mapped to the first round. The input values ​​for each evaluation dimension adopt the same unified and standardized criteria as in step 400; For the first The trait category in the first Mean parameters across each evaluation dimension; For the first The trait category in the first Variance parameters on each evaluation dimension; To prevent the stability constant from having a denominator of 0; Sum index for categories; The number of potential trait categories; To assess the number of dimensions, the above calculations show that both interim flight performance and newly added assessment data are used to update the classification probability.

[0038] In this embodiment, after obtaining the updated classification probabilities, the potential profile type or feature intensity of the trainee is adjusted based on the classification probabilities to dynamically update the flight potential profile and form a phased update result. The potential profile type is determined by the prototype category corresponding to the highest classification probability, and the feature intensity is updated by combining the classification probability weighting with the historical profile smoothing. Specifically, it can be expressed as follows: in, For the first The student in the first The confidence score of the updated portrait after each round of updates is used to characterize the degree of certainty of the classification results in that round and serves as the portrait feature strength update coefficient. For the first The student in the first The first round of updates Each evaluation dimension profile feature strength; For the first The student in the first The first round of updates The strength of the portrait features in each evaluation dimension is calculated from the prototype core feature vector and the initial classification probability in step 400. For the first The prototype in the first Core feature vector components across each evaluation dimension; For the first The student in the first The probability of category assignment after each round of updates; This represents the number of potential trait categories. The phased update results include at least the potential profile type, the profile feature strength of each assessment dimension, the profile update confidence level, and the direction of change relative to the previous round.

[0039] In this embodiment, to feed the data generated during the dynamic update process back into the flight potential classification model and perform iterative optimization, this embodiment uses a method of updating the prototype's interim flight performance benchmark value and the prototype's core feature vector after screening high-confidence samples for incremental optimization. Specifically, the following formula can be used: in, For the first The student in the first During the first update, the first... Feedback weights for each prototype category are used to filter and weight the contribution of high-confidence samples to the model parameters; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. To provide feedback on the confidence threshold, it is preferable to choose a value between 0.60 and 0.90, with 0.75 being an example. For the first The prototype category in the first The benchmark values ​​for the staged flight performance after rounds of iterative optimization; For the first The prototype category in the first The 1st round of iterative optimization Core feature vector components of each evaluation dimension; The number of student samples participating in the current round of iteration optimization; For the first The student in the first The phased flight performance of the wheel; For the first The student in the first The first round of updates Each evaluation dimension profile feature strength; , and The meaning is consistent with the previous text. In this embodiment, samples that have reached the feedback confidence threshold participate in parameter updates to reduce the impact of low-confidence samples on prototype parameter drift and output phased update results.

[0040] In this embodiment, the flight potential profile in step 500 refers to the phased structured representation result formed for a single trainee, including at least the potential profile type, the profile feature strength of each assessment dimension, classification confidence, profile update confidence, and assessment round identifier. The potential profile type refers to the prototype category identifier determined based on the category belonging probability output by the flight potential classification model, used to represent the trainee's current flight potential category. The profile feature strength refers to the vectorized representation value of the trainee on each assessment dimension, used to describe the relative strength of the flight ability dimension and the non-ability factor dimension. The classification confidence refers to the probability value or an index calculated from the probability value used to characterize the reliability of the classification result. The profile update confidence refers to the stability of the current round's dynamic update result and controls the fusion of old and new profiles. The quantitative indicators of amplitude; the stage evidence item refers to the category support quantity formed by the historical classification probability, stage flight performance and newly added evaluation data, which is used to calculate the updated category classification probability; feedback refers to the process of using the high-confidence stage results formed during the dynamic update process to update the parameters of the flight potential classification model; feedback weight refers to the quantitative weight that characterizes the contribution of the sample to the update of the prototype parameters during iterative optimization; high-confidence sample refers to the sample whose profile update confidence reaches the preset feedback confidence threshold, and its function is to reduce the impact of boundary samples on the drift of prototype parameters; stage update result refers to the data set output by step 500 that can be called by step 600, which includes at least the updated potential profile type, the profile feature strength of each evaluation dimension, the profile update confidence and the change direction information relative to the previous round.

[0041] Optionally, in this embodiment, step 500 corresponds to the process of generating and initially applying a flight potential profile. Specifically, based on the flight potential classification model formed in step 400, this embodiment performs quantitative modeling and classification of trainees, and generates a flight potential profile according to the trainee's prototype category, descriptive labels, and core feature vectors. The flight potential profile is used to characterize the trainee's potential profile type and the intensity of features in each dimension under the current assessment round. Further, this embodiment correlates the quantitative modeling and classification results with subsequent flight training data to achieve digital prediction of flight potential. The data correlation includes at least establishing a round-based correspondence between the trainee's profile type, classification probability, core feature vectors, and the stage-based flight scores, real flight performance, and newly added assessment data generated in subsequent training phases. This allows this embodiment to not only output static classification results but also provide structured input for continuous tracking, dynamic updates, and prediction model construction in subsequent training phases. Through the above methods, this embodiment transforms the category recognition results obtained from potential profile analysis into a flight potential profile that can be directly called in flight training scenarios, realizing data integration from statistical modeling and classification to training application.

[0042] In this embodiment, step 500 further corresponds to the dynamic update of the flight potential profile and the iterative optimization process of the model, and is connected with step 600. Specifically, this embodiment treats each trainee as a dynamic entity. When a trainee enters the subsequent learning and training stage and obtains real flight performance, the real flight performance, the stage flight score, and the newly added evaluation data are input into the system. This embodiment not only updates the trainee's current performance record, but also dynamically updates the potential profile type or feature strength of the trainee through an algorithm. The algorithm preferably includes a classification probability adjustment based on Bayesian update to correct the trainee's probability of belonging to each prototype category based on new evidence, and adjusts the potential profile type or feature strength of each dimension accordingly. Meanwhile, this embodiment continuously accumulates the data generated during the dynamic update process and feeds the data back into the latent classification model, enabling the latent classification model to be iteratively optimized in subsequent rounds, thereby gradually improving the stability and adaptability of the classification and update results. This forms a closed-loop process of data collection, evaluation and calibration, profile generation and updating, training intervention, and new data generation. The phased update results output in step 500 can be further used as the input basis for the future flight performance prediction model in step 600, so as to support the continuous prediction and correction of the trainee's performance level, risk tendency, and adaptability in future specific flight missions or training phases.

[0043] Furthermore, in this embodiment, step 600 constructs a predictive model for future flight performance based on the flight potential profile and phased update results output in step 500, and uses it to predict the trainee's performance level, risk propensity, and adaptability in a specific future flight mission or training phase. Preferably, the predictive model adopts a structure of a shared representation layer and a multi-output prediction layer, wherein the shared representation layer is used to integrate historical flight data, current capability status, and the evolution trend of non-capability factors, and the multi-output prediction layer is used to output predicted values ​​for performance level, risk propensity, and adaptability, respectively; wherein, historical flight data preferably forms time-series statistical features according to a preset historical window, the current capability status preferably consists of the feature intensity of the flight potential profile in step 500, and the evolution trend of non-capability factors preferably consists of the change in the relevant assessment dimensions of non-capability factors in consecutive rounds. The predictive output can be expressed by the following formula: in, For the first The student in the first The fusion input vector used for prediction includes time-series statistical features formed from historical flight data, current capability status, and evolution trends of non-capability factors; For the first The student in the first Shared representation vector of the wheel; For nonlinear mapping functions, hyperbolic tangent function, rectified linear function or equivalent monotonic nonlinear function are preferred; For the first The round prediction model shares the representation layer parameter matrix; For the first Round prediction models share the representation layer bias vector; For the first The first student faces the future. The predicted output vector of the wheel is preferably defined as follows: The first component corresponds to the predicted performance level, the second component corresponds to the predicted risk propensity, and the third component corresponds to the predicted adaptability. For the first Round prediction model with multiple output prediction layer parameter matrices; For the first The round prediction model outputs multiple prediction layer bias vectors; The prediction step size represents the predicted interval for a specific future flight mission or training phase relative to the current round, preferably an integer between 1 and 5 rounds.

[0044] In this embodiment, after obtaining the newly added flight performance data for each round, the prediction result for that round is compared with the actual flight performance data, and the prediction model output is dynamically corrected based on the deviation through a recursive calibration mechanism. Preferably, the newly added flight performance data includes at least the actual performance level, actual risk event characterization results or risk scores, and actual adaptability assessment results, each corresponding one-to-one with the aforementioned three prediction output components. Preferably, the prediction deviation is calculated first, then the output calibration item is updated based on the prediction deviation, and the updated output calibration item is superimposed on the prediction results of subsequent rounds. Specifically, the following formula can be used: in, For the first The student in the first The newly added actual flight performance vector is preferably defined as follows: And corresponding to the actual observed values ​​of performance level, risk propensity and adaptability, respectively; For the first The student in the first The prediction deviation vector of the wheel; For the first The output calibration term vector is used to characterize the amount of systematic bias compensation of the prediction model for the three output targets at the current stage; For the first The preferred prior uncertainty of the calibration round is obtained from the statistical analysis of the residuals of the previous calibration round. For the first The newly added deviation dispersion is preferably obtained from the statistical analysis of the prediction deviation of the current round or the most recent consecutive rounds; For the first Round recursive calibration gain is used to control the update magnitude of the output calibration term by the prediction bias of the current round; The aforementioned stability constant; For the first The future number of students after dynamic correction The round predicts the output vector. When using Bayesian update, it can be... Treat it as a posterior correction term, and They correspond to prior uncertainty and observation uncertainty respectively, and are updated according to the same data organization caliber.

[0045] In this embodiment, to avoid a continuous decline in prediction performance due to changes in sample distribution or target shift in consecutive rounds, a retraining trigger rule based on validation data statistics and consecutive exceedance conditions is set. Preferably, the prediction error continuously exceeding a preset threshold is determined by a combination of normalized comprehensive error and consecutive round counts, to simultaneously consider the dimensional differences among the three output targets: performance level, risk propensity, and adaptability. Preferably, the normalized comprehensive error, retraining threshold, and consecutive exceedance counts can be calculated using the following formula: in, For the first The student in the first The normalized composite error of the round is used as the basic error metric for determining retraining triggers; , and The first The student in the first The prediction bias components of the corresponding performance level, risk propensity, and adaptability; , and The error normalization scaling parameters for the three prediction targets are preferably determined by the standard deviation, quantile difference, or upper limit of the business allowable error of the corresponding targets in the validation set during the training phase, so as to ensure the comparability of the three targets in the comprehensive error calculation. The preferred threshold for triggering model retraining is the average comprehensive error normalized from the validation set. Standard deviation of normalized composite error with validation set It is determined that the mean plus three times the standard deviation is used to improve the robustness of the threshold setting and reduce the probability of occasional anomalies. As of the date The wheel is of length The number of out-of-limit counts within a continuous window; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Index for summing rounds within the window; To ensure continuous determination of the window length, it is preferable to choose an integer between 3 and 10 rounds. Preferably, when... Time-triggered model retraining, i.e., continuous The normalized synthesis error of each round exceeded the retraining trigger threshold.

[0046] In this embodiment, after triggering model retraining, the training samples are reconstructed and the prediction model parameters are updated based on the flight potential profile, phased update results, historical flight data, and corresponding actual flight performance data to obtain a dynamically corrected prediction model; wherein, the input of the training samples still uses the same method as described above. The data organization standards remain consistent, and the output of the training samples still uses the same methods as described above. A consistent three-target output caliber is used to maintain consistency between the training and application phases. Preferably, the prediction model training should expose and fix at least the following necessary training information: prediction step size. The range and selection principles of the values, the historical window length and window statistics method, the method of dividing the training set and validation set, the termination condition, the model parameter initialization method, and the sample update range and sample inclusion rules during retraining; among them, the retraining samples preferably include the newly added samples from the most recent consecutive rounds and historical high-confidence samples, the number of recent samples is preferably an integer between 20 and 200, and the high-confidence samples can be obtained by filtering the feedback confidence threshold in step 500. The performance level is preferably quantified by the flight mission completion quality or stage assessment score, the risk tendency is preferably quantified by the probability of occurrence of risk events, risk score or warning level mapping value, and the adaptability is preferably represented by a quantitative indicator of the performance recovery speed, stability or standard attainment under new tasks, new rules or new operational load conditions; the above three types of outputs use the same definition and dimensionality in the training data construction and online prediction stages to ensure the consistency of the data closed loop between prediction results, deviation calculation, dynamic correction and retraining trigger.

[0047] In this embodiment, the prediction model in step 600 refers to a multi-objective prediction model used to output performance level, risk propensity, and adaptability for a specific future flight mission or training phase. The shared representation layer refers to a parameter layer that fuses and encodes historical flight data, current capability status, and the evolution trend of non-capability factors, used to extract representation information shared by multiple prediction objectives. The multi-output prediction layer refers to a parameter layer that generates predicted values ​​for performance level, risk propensity, and adaptability based on the shared representation. The historical window refers to the range of consecutive rounds used to construct the time-series statistical characteristics of historical flight data. The evolution trend of non-capability factors refers to the trend quantification result formed by the direction and magnitude of change of relevant assessment dimensions of non-capability factors in consecutive rounds. The output calibration term refers to the vectorized correction amount used to compensate for systematic biases in the prediction model. The recursive calibration gain refers to the gain calculated based on the calibration prior uncertainty and the new bias. The update amplitude coefficient obtained from the divergence calculation; calibration prior uncertainty refers to the statistic characterizing the reliability of existing calibration items; new deviation dispersion refers to the statistic characterizing the fluctuation of prediction deviation in the current round or several recent consecutive rounds; normalized comprehensive error refers to the comprehensive error index obtained by normalizing the prediction deviations of the three output targets of performance level, risk tendency, and adaptability according to the corresponding error normalization scale parameters, which is used to unify the different target dimensions and serve as the basis for retraining trigger judgment; error normalization scale parameters refer to the standard deviation, quantile difference, or business allowable error upper limit corresponding to the three output targets respectively; retraining trigger threshold refers to the error upper limit judgment value determined based on the validation data statistics; continuous judgment window length refers to the window range used to statistically analyze the rounds in which the error continues to exceed the limit; the above terms are used together to form the closed-loop judgment logic of prediction, deviation calculation, dynamic correction, and retraining trigger in step 600.

[0048] As an example, in this embodiment, to further define the combination of the prediction model and the flight training scenario, this embodiment constructs the prediction samples in a supervised manner according to the rounds: taking the first round as an example... The round time point is used as the prediction input time point, utilizing the student up to the [number]th round. The input vector consists of the flight potential profile of the wheel, the results of phased updates, the time-series statistical characteristics of historical flight data, the current capability status, and the evolution trend of non-capability factors. And based on the student's future... The actual flight performance data corresponding to a specific flight mission or training phase constitutes the supervision label. This ensures a one-to-one correspondence between input and output in terms of trainee identification, round index, task stage, and prediction step size. The performance level label is preferably formed from stage assessment scores, task completion quality scores, or key operation pass rates; the risk propensity label is preferably formed from risk event occurrence records, risk scores, or warning level mapping values; and the adaptability label is preferably formed from the time to achieve the target under new task conditions, performance recovery speed, number of consecutive stable target achievement rounds, or fluctuation amplitude indicators. The same definition, unit processing rules, and missing value handling rules are used in both the training and online prediction phases. For different task types or training stages, this embodiment preferably adds task type identifiers and training stage identifiers as input features, enabling the prediction model to learn the differential association between input features and the three prediction targets under different flight scenarios. When applying the model output, the first component of the prediction output vector is used to generate the performance level prediction result, the second component is used to generate the risk propensity prediction result, and the third component is used to generate the adaptability prediction result. These are then compared with actual observations of the same caliber for deviation calculation, dynamic correction, and retraining trigger determination, forming a closed loop of prediction, calibration, and retraining in step 600.

[0049] Optionally, in this embodiment, step 600 corresponds to the prediction process of future flights based on the dynamic profile. This embodiment, based on the formation of a dynamically updated flight potential profile, constructs a prediction model for future flight performance. This model is used to assess the trainee's performance level, risk tolerance, and adaptability in specific training phases or actual flight missions in the future, and serves as a forward-looking and adaptive prediction engine. The prediction model is not a simple extrapolation of historical performance, but is constructed based on a multi-dimensional fusion prediction framework. Its core inputs include at least static and dynamic profile data, task and environmental characteristics, and time and sequence factors. The static and dynamic profile data includes at least personal potential classification labels, historical flight performance sequences, skill mastery curves, baselines of non-capability factors, and trends in non-capability factor changes. The task and environmental characteristics include at least task complexity, expected environmental stress level, and crew configuration requirements obtained by characterizing future flight missions. Task complexity can correspond to task types such as instrument flight, aerobatic maneuvers, and emergency handling. The time and sequence factors include at least training time intervals and the sequence dependency of historical performance, used to characterize the impact of previous rounds or several previous rounds of performance on subsequent rounds. Through the above-described input organization method, this embodiment will jointly model the flight potential profile and phased update results from step 500 with specific flight training scenarios, mission conditions, and time evolution processes.

[0050] In this embodiment, the prediction method in step 600 adopts a hybrid modeling approach to balance the ability to analyze long-term development trajectories with the ability to predict continuous performance indicators. Specifically, on the one hand, based on the classification results obtained from the potential profile analysis in step 400, this embodiment uses state transition analysis and trajectory modeling methods to model the transition probabilities between different potential states of trainees. Markov chain models or hidden Markov models are preferred to predict the trainee's long-term capability development path and output the probability of transitioning from the current potential state to the target stable state. For example, this can be used to assess the probability of a state with strong technical capabilities but relatively insufficient adaptability evolving into a fully stable state. On the other hand, this embodiment performs multivariate time series prediction for continuous key flight performance indicators. These key flight performance indicators include at least landing deviation, navigation accuracy, and program execution fluency. Long short-term memory networks, Transformer time series models, or dynamic structural equation models are preferred to capture the nonlinear patterns of the key flight performance indicators over time, thereby predicting the performance range at future time points or in future training phases. By combining state transition analysis with multivariate time series forecasting, this embodiment simultaneously predicts development paths, stage performance, and risk change trends within the same forecasting model system.

[0051] In this embodiment, step 600 outputs and presents the prediction results in a multi-dimensional manner using decision support. Specifically, this embodiment outputs the prediction results as one or more of the following: a visualized trend curve, a risk warning level or matrix, an adaptability interval, and a development path diagram. The visualized trend curve displays the predicted values ​​and confidence intervals of core skill indicators over the next 1 to 3 training phases; the risk warning level or matrix is ​​used to classify potential risks and identify risk types, with the risk level preferably including low, medium, and high, and the risk types preferably including skill decay, decision-making ability degradation under stress, and weaknesses in specific subjects, to indicate the key directions for training intervention; the adaptability interval and task matching suggestions provide suitable task types for trainees, appropriate partner characteristics, and scenarios that need to be avoided or prepared for more; the development path diagram visualizes the trainee's two to three most likely ability development trajectories to support the development of personalized advanced training plans. The above output results correspond to the three prediction targets of performance level, risk propensity, and adaptability in step 600, and are consistent with the collection criteria of subsequent actual flight performance data to facilitate subsequent deviation calculation and dynamic correction.

[0052] In this embodiment, step 600 further includes a dynamic calibration and update mechanism for the prediction model to form a long-term closed-loop feedback process. Specifically, this embodiment adopts a rolling prediction method. After obtaining a new round of actual flight results and new evaluation data, the prediction results of the corresponding round are compared with the actual values, the prediction deviation is calculated, and the prediction model parameters are dynamically adjusted based on a Bayesian update mechanism or a recursive calibration mechanism. The prediction model parameters preferably include state transition probabilities, output calibration terms, or parameters used for predicting output mapping, so that subsequent prediction results are more consistent with the student's recent actual performance. Furthermore, this embodiment continuously monitors the prediction error and introduces an uncertainty decay or uncertainty adjustment mechanism for long-term predictions. When the prediction error continues to exceed a preset threshold, model retraining is triggered or the administrator is prompted to perform manual diagnosis to avoid long-term model drift leading to prediction distortion. In addition, this embodiment can also use group samples for periodic learning under the premise of privacy protection, that is, use the prediction results and actual performance data of all students to periodically iterate and optimize the general prediction model to improve the system's initial prediction ability for new students. Thus, this embodiment forms a continuous cycle of evaluation, prediction, verification, and calibration, so that step 600 is not only used to output future performance prediction results, but also to support refined and personalized flight talent training and safety risk management.

[0053] The beneficial effects of this invention are as follows: (1) This invention establishes a mapping relationship between expert evaluation scores and key objective flight parameters, and performs parametric calibration of the original subjective evaluation according to preset rules, transforming the scoring correction process that previously relied on individual experience into a reusable and computable calibration process. On the one hand, it can constrain subjective bias, scale drift, and individual strictness differences in dual-expert scoring, improving the consistency between the calibrated expert evaluation and actual operational performance; on the other hand, it can precipitate the calibration experience formed by experts over a long period of time into model rules for subsequent evaluation of new samples, thereby improving the stability, comparability, and consistency of repeated implementation of the evaluation results, and providing higher quality input data for subsequent potential classification and predictive modeling.

[0054] (2) This invention gathers multi-dimensional assessment data by combining calibrated expert evaluations, aggregated objective parameters, and psychological test data. Based on potential profile analysis, it forms a flight potential classification model, realizing a shift from judging single scores to recognizing overall patterns. Compared to methods that only evaluate a single ability indicator, this invention can identify the trainee's potential trait categories from the combination of flight ability and non-ability factor dimensions, and generate descriptive labels and core feature vectors. This allows the assessment results to reflect not only the trainee's current performance but also their potential strengths, weaknesses, and training direction. As a result, the assessment conclusions are closer to the actual composition of pilot competence and are more conducive to hierarchical management, targeted intervention, and personalized training in the training process.

[0055] (3) After completing the flight potential classification and profile generation, this invention further inputs the phased flight performance and newly added assessment data into the flight potential classification model to dynamically update the flight potential profile. Based on the dynamic update results, a future flight performance prediction model is constructed and corrected, thus forming a continuously evolving assessment closed loop. The output of this technical solution is not a one-time static result, but a dynamic profile and prediction result that can be continuously updated with the training process. It can reflect the consolidation, fluctuation, or change trend of the trainee's potential type and continuously correct the prediction deviation in combination with newly added flight performance data. With the accumulation of samples and model iteration, a mutually reinforcing relationship is formed between assessment calibration, potential classification, and future performance prediction, which helps to improve assessment accuracy, prediction timeliness, and training decision support capabilities.

[0056] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0057] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting flight potential based on expert experience and dynamic profile generation technology, characterized in that, include: Construct a multi-dimensional evaluation table that includes flight capability dimensions and non-capability factor dimensions; Based on the aforementioned multidimensional evaluation table, two experts independently score the performance of the same student in the same subject, forming two original subjective evaluations. Establish a mapping relationship between expert evaluation scores and key objective flight parameters, and perform parameterized calibration on the two original subjective evaluations according to preset rules to obtain calibrated expert evaluations; The calibration-based expert evaluation, objective parameter aggregation indicators, and psychological test data are combined to form multi-dimensional evaluation data. Based on the multi-dimensional evaluation data, potential profile analysis is performed to form a flight potential classification model. The flight potential classification model is used to classify the flight potential of the trainees and generate a flight potential profile. The phased flight performance and new evaluation data are input into the flight potential classification model to dynamically update the flight potential profile and obtain the phased update results. Based on the flight potential profile and the phased update results, a predictive model for future flight performance is constructed. The predictive model is dynamically corrected according to the newly added flight performance data in each round. Based on the dynamically corrected predictive model, the trainee's performance level, risk propensity, and adaptability in future specific flight missions or training phases are predicted.

2. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 1, characterized in that, Construct a multi-dimensional evaluation table that includes flight capability dimensions and non-capability factors, including: Set the evaluation items for the flight capability dimension and the evaluation items for the non-capability factor dimension; For the evaluation items of the flight capability dimension and the evaluation items of the non-capability factor dimension, corresponding expert observation clues are configured respectively; The evaluation items are configured with corresponding expert observation clues to form the multi-dimensional evaluation table.

3. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 2, characterized in that, The evaluation items for the flight capability dimension and the evaluation items for the non-capability factor dimension are set, including: Evaluation items are set in the flight capability dimension, including comprehension and acceptance, memory ability, control ability, attention ability, spatial orientation, adaptability and situational awareness; Evaluation items for emotional control, willpower, motivation, and personality are set in the non-ability factor dimension; The expert observation cues are configured according to the evaluation items so that the two experts can score independently based on the multi-dimensional evaluation table.

4. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 1, characterized in that, Based on the aforementioned multidimensional evaluation table, two experts independently score the performance of the same student in the same subject, resulting in two original subjective assessments, including: Based on the aforementioned multidimensional evaluation table, the two experts independently score the performance of the same student in the same subject based on the same flight data and performance records, forming the two original subjective evaluations. Correlation analysis and difference testing were performed on the two original subjective assessments. Calculate the consistency index for each evaluation dimension; When the score difference of a key project exceeds a preset threshold, a consistency warning is triggered; In response to the consistency alert, a cross-check was performed on the two original subjective assessments.

5. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 1, characterized in that, Establish a mapping relationship between expert evaluation scores and key objective flight parameters, and perform parameterized calibration on the two original subjective evaluations according to preset rules to obtain calibrated expert evaluations, including: Establish a mapping relationship between the expert evaluation scores and the key objective flight parameters; Based on the mapping relationship, determine whether the two original subjective assessments are consistent with the operational performance reflected by the key objective flight parameters; When the judgment result indicates that there is a significant deviation, the two original subjective evaluations are parametrically calibrated according to the preset rules to obtain the calibrated expert evaluation. If the judgment result is that there is no significant deviation, the two original subjective assessments are determined as the calibrated expert assessments.

6. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 5, characterized in that, When the judgment result indicates a significant deviation, the two original subjective evaluations are parametrically calibrated according to the preset rules to obtain the calibrated expert evaluation, including: When one of the original subjective assessments is significantly inconsistent with the key objective flight parameters, and the other original subjective assessment is highly consistent with the key objective flight parameters, the original subjective assessment with high consistency with the key objective flight parameters shall be adopted first, and the abnormal scores shall be weighted and adjusted to obtain the calibrated expert assessment. When both original subjective assessments deviate from the key objective flight parameters and the deviations are in the same direction, the two original subjective assessments are synchronously calibrated according to the parameter rules to obtain the calibrated expert assessment. When both original subjective assessments are highly consistent with the key objective flight parameters and there are reasonable differences between them, the reasonable differences are retained. Based on the reasonable differences, an interval score or a composite score is generated as the post-calibration expert evaluation.

7. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 5, characterized in that, The two original subjective evaluations are parametrically calibrated according to the preset rules to obtain the calibrated expert evaluation, which includes: Train a regression model using historical data; The set of objective parameters and the original expert scores are used as inputs to the regression model; The standard score, after arbitration by two authoritative experts, is used as the output of the regression model. The regression model is used to learn a function to correct for the bias; The new evaluation is parametrically calibrated based on the learned correction bias function to obtain the calibrated expert evaluation.

8. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 1, characterized in that, The calibrated expert evaluations, aggregated objective parameters, and psychological test data are combined to form multi-dimensional evaluation data. Based on this multi-dimensional evaluation data, potential profile analysis is performed to form a flight potential classification model, including: The multi-dimensional assessment data is formed by combining the calibrated expert evaluation, the aggregated objective parameter indicators, and the psychological test data. Based on the aforementioned multidimensional assessment data, potential profile analysis is applied to perform statistical modeling to identify potential trait categories; Based on the aforementioned trait categories, trainees are prototyped according to the flight capability dimension and the non-capability factor dimension; Descriptive labels and core feature vectors are generated for each of the obtained prototypes; The flight potential classification model is formed based on the prototype, the descriptive labels, and the core feature vectors.

9. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 1, characterized in that, The flight potential classification model is used to classify the trainees' flight potential and generate flight potential profiles. Interim flight performance and new assessment data are input into the flight potential classification model to dynamically update the flight potential profiles, resulting in interim update results, including: The flight potential classification model is used to classify the flight potential of the trainees to obtain the corresponding classification results for the trainees; A flight potential profile of the trainee is generated based on the classification results; Input the interim flight results and the newly added assessment data into the flight potential classification model; The potential profile type or feature strength of the trainee is adjusted based on the classification probability in order to dynamically update the flight potential profile. The data generated during the dynamic update process is fed back into the flight potential classification model to iteratively optimize the model and output the phased update results.

10. The flight potential prediction method based on expert experience and dynamic profile generation technology according to claim 1, characterized in that, Based on the flight potential profile and the phased update results, a predictive model for future flight performance is constructed. The predictive model is dynamically revised based on the newly added flight performance data in each round. Based on the dynamically revised predictive model, the trainee's performance level, risk tolerance, and adaptability in future specific flight missions or training phases are predicted, including: Based on the flight potential profile and the phased update results, the prediction model is constructed by integrating historical flight data, current capability status, and the evolution trend of non-capability factors. After obtaining the new flight performance data for each round, the deviation between the predicted value and the actual value is compared with the prediction results and actual flight performance data for the corresponding round. Based on the aforementioned deviation, the prediction model is dynamically corrected using a Bayesian update or recursive calibration mechanism. When the prediction error continues to exceed a preset threshold, the model is retrained. The prediction model, after dynamic correction, predicts the trainee's performance level, risk propensity, and adaptability in future specific flight missions or training phases.

Images