A flight training-oriented behavior feature extraction and manipulation mode analysis method

Abstract

The application relates to the technical field of flight training evaluation, and discloses a flight training-oriented behavior feature extraction and manipulation mode analysis method, which comprises the following steps: synchronously collecting flight parameters, hand and instrument gaze videos and loading a standard flight path template; calculating a phase difference of a manipulation instruction relative to the template to represent a perception advance; calculating a correlation between gaze switching and manipulation frequency to determine a coherence coefficient; calculating a cognitive load weight coefficient according to the phase difference and the coherence coefficient; mapping multi-modal data to a behavior embedding space to generate a behavior representation vector; and using the weight coefficient to correct a change trend of the vector to generate a manipulation stability index. The application quantifies cognitive feedback efficiency behind manipulation, removes a false stability component generated by compensatory manipulation, and solves the problem of identifying that a standard manipulation result covers up a manipulation logic deviation.

Description

Technical Field

[0001] This invention relates to a method for extracting behavioral features and analyzing control patterns for flight training, belonging to the field of flight training assessment technology. Background Technology

[0002] Current flight training assessment systems use recording systems to collect flight trajectory and attitude data, compare control results to determine performance, and regard flight trajectory compliance as skill mastery. In high-dynamic training, compliance of results often masks logical deviations. Trainees achieve their goals through high-frequency disordered compensatory corrections. Existing multimodal analysis of execution action records and feature classification cannot quantify the feedback efficiency of the perception-decision-execution link, and it is difficult to distinguish between skilled control and compensatory control.

[0003] To address these shortcomings, the industry has attempted to introduce deep learning networks to perform feature space mapping. However, relying solely on data-driven approaches can easily overlook the inherent physical constraints of flight control. For example, Chinese invention patent CN120911788A discloses a method, system, device, and storage medium for dynamic assessment of pilots based on a TEM model. Although the solution introduces multimodal data streams and utilizes a dual-channel long short-term memory network (LSTM) for deep feature extraction and decision tracing, its core logic still focuses on the classification and statistics of macroscopic threat-error results, without deeply analyzing the underlying mechanisms of control behavior. Such methods tend to establish a black-box mapping between input data and assessment conclusions, ignoring the inherent physical constraints of flight control. In particular, they lack quantitative analysis of the frequency domain coupling relationship and phase lead quantity between visual perception and physical actions. This makes it difficult for the algorithm to remove random fluctuation noise caused by physiological tension or blind operation of trainees when facing highly dynamic and complex conditions. It cannot identify the pseudo-stable state that is temporarily upheld by high-frequency corrections, even though the flight path data is temporarily upheld. Furthermore, it cannot solve the problem of quantifying perception lead quantity, leading to the risk of false alarms in the assessment conclusions under high-speed dynamic environments.

[0004] Therefore, how to quantify the efficiency of manipulating cognitive feedback and construct a manipulation pattern profile is the technical problem that this invention aims to solve. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for extracting behavioral features and analyzing control patterns for flight training, comprising the following steps:

[0006] Step S1: Simultaneously collect flight quick access recorder parameters, hand movement video data, and instrument gaze area video data during flight training, and load the corresponding standard trajectory response template according to the flight mission stage.

[0007] Step S2: Analyze the control command sequence in the flight quick access recorder parameters, and perform phase comparison calculation between the control command sequence and the standard track response template to generate the command response phase difference that characterizes the trainee's control perception lead.

[0008] Step S3: Use the action recognition model to extract the motion trajectory function from the hand action video data, and use the visual attention model to determine the temporal probability distribution of the gaze point in the instrument gaze area video data.

[0009] Step S4: Calculate the cross-correlation degree between the temporal probability distribution of the gaze point and the spectral characteristics of the motion trajectory function, and determine the coherence coefficient representing the degree of coupling between the visual search logic and the manipulation execution action based on the cross-correlation degree;

[0010] Step S5: Calculate the cognitive load weighting coefficient based on the command response phase difference and coherence coefficient, and convert the flight rapid access recorder parameters and video data into a preset behavior embedding space through a nonlinear mapping algorithm to generate a behavior representation vector that represents the trainee's comprehensive operational state.

[0011] Step S6: Perform a weighted correction operation on the temporal change trend of the behavior representation vector using the cognitive load weight coefficient to generate the manipulation stability index after removing the compensatory manipulation component.

[0012] Preferably, when generating the command response phase difference, the following operations are performed: second derivative operations are performed on the target control command sequence contained in the standard track response template and the acquired actual control command sequence to identify command jump points where the command rate of change changes abruptly; characteristic timestamps of the actual control command sequence at the command jump points are extracted, and the time offset of the characteristic timestamp relative to the corresponding jump point in the target control command sequence is calculated; the average offset value of multiple command jump points is used to determine the trainee's perception lead in the current flight mission phase.

[0013] Preferably, step S4 specifically includes: defining the gaze switching frequency in the probability distribution time series as the driving signal, defining the manipulation frequency corresponding to the motion trajectory function as the response signal; determining whether the manipulation behavior is controlled by the visual search logic by analyzing the overlap of the energy distribution of the driving signal and the response signal in the frequency domain, and defining the value of the overlap of the energy distribution as the coherence coefficient.

[0014] Preferably, when calculating the cognitive load weighting coefficient in step S5, the following quantification rules are followed: ,in, For cognitive load weighting coefficient, For the command response phase difference, The coherence coefficient, as well as These are the preset sensitivity adjustment parameters.

[0015] Preferably, step S6 specifically includes: real-time monitoring of the numerical change of the coherence coefficient; if the coherence coefficient is lower than the preset coupling threshold and the command response phase difference shows a non-linear fluctuation trend, then it is determined that the trainee's current stage of manipulation behavior belongs to blind compensatory manipulation.

[0016] Preferably, in cases where the manipulation is determined to be blind compensatory manipulation, the rate of change of the behavioral representation vector is doubled using the cognitive load weight coefficient to reduce the spurious stability weights generated by blind compensation in the manipulation stability index.

[0017] Preferably, after generating the control stability index, the following operations are also included: acquiring voice audio data during flight training and extracting acoustic indicators representing psychological stress from the voice audio data; using the acoustic indicators to calculate the confidence level of the control stability index and attaching a corresponding confidence level label to the control stability index.

[0018] Preferably, the steps for extracting acoustic indicators include: analyzing the pitch change rate, pause length, and fundamental frequency fluctuation amplitude of the speech audio data; if any parameter of the pitch change rate, pause length, or fundamental frequency fluctuation amplitude exceeds a preset stress threshold, the weight ratio of the cognitive load weight coefficient in the correction calculation is increased.

[0019] Preferably, it also includes: analyzing the long-sequence characteristic trend of the manipulation stability index during the training process, and classifying the manipulation behavior of trainees into predictive, lagging, or over-controlling manipulation styles.

[0020] Preferably, in step S1, the acquired flight quick access recorder parameters, hand movement video data, instrument gaze area video data, and voice audio data are time-aligned using a unified clock synchronization mechanism, so that the manipulation behavior at any given time corresponds to a unique joint feature vector.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. In the extraction of behavioral features in flight training, a standard flight path response template is loaded according to the semantic structure of the flight mission. The phase difference between the response of the target control command and the actual control command is calculated. The cognitive feedback efficiency behind the control behavior is quantified. The pilot's perception lead time of the flight status is identified. The phase characteristics of the command response are compared. The predictive skillful control and the delayed compensatory correction are distinguished. This solves the problem that traditional evaluation methods cannot identify the correctness of control logic based on the standard of control results.

[0023] 2. Based on the probability distribution change sequence of the gaze area and the spectral characteristics of the hand movement trajectory function, the coherence is determined, the cross-correlation coherence coefficient between visual search and physical execution is calculated, and a closed loop of logical association between visual perception and action execution is constructed. When the coherence coefficient is lower than the preset threshold and the manipulation frequency is abnormally increased, the current stage is determined to be blind compensatory manipulation. This identifies whether the pilot has established a stable manipulation prediction ability, explains that the root cause of the operation deviation is perception lag or visual search logic confusion, and provides causal logic support data for personalized flight teaching guidance.

[0024] 3. Convert flight parameters, video motion features, and audio acoustic indicators into behavioral embedding space representation vectors to generate a behavioral representation that characterizes the pilot's comprehensive behavioral state. Use the cognitive load weighting coefficient generated by weighting the response phase difference and coherence coefficient to correct behavioral trends. Use mission intent as a frequency anchor point to filter out random fluctuation noise caused by the trainee's physiological tension or blind operation, remove false stability from the control mode profile, ensure that the control style classification is consistent with skill proficiency, and enhance the anti-interference ability of behavioral feature extraction in high-dynamic flight scenarios. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the behavioral feature extraction and stability evaluation process of the present invention;

[0026] Figure 2 This is a schematic diagram illustrating the temporal synchronous evolution relationship of the multimodal features of the present invention;

[0027] Figure 3 This is a schematic diagram of the multi-source perception and intelligent analysis architecture of the present invention. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0029] This invention provides a method for behavioral feature extraction and control pattern analysis in flight training. By integrating flight parameters, hand gesture video data, and instrument gaze area video data, a synchronous analysis architecture for multimodal data streams is constructed. The method calculates the coherence between the trainee's perceived lead and execution logic during control operations, and generates a control stability index after removing compensatory control components. The method includes steps such as synchronous data acquisition and preprocessing, loading of standard trajectory response templates, phase-locked control entropy measurement, coherence analysis of gaze and gestures, calculation of cognitive load weighting coefficients, and control pattern recognition based on behavioral embedding space. The system performs synchronous data acquisition and preprocessing. Through a unified clock synchronization mechanism, the system connects to the onboard data bus of the flight simulator or training aircraft and external monitoring equipment, synchronously acquiring flight fast access recorder parameters, hand gesture video data, instrument gaze area video data, and audio data. The system sets a global sampling frequency such as... The system uses the Network Time Protocol (NTP) to synchronize the time of each acquisition node, ensuring that all data frames are marked with a unified timestamp. For data sources with inconsistent frequencies, the system uses a linear interpolation algorithm to upsample low-frequency data to the global frequency, thereby generating a time-aligned joint feature vector. This vector contains the current time. The flight attitude parameters, joystick position commands, thrust stick position commands, hand spatial coordinates, gaze point screen coordinates, and audio fundamental frequency characteristics.

[0030] The system performs mission intent benchmark alignment. Based on the semantic structure of the current flight mission, such as the glide phase in a five-way flight, the system loads the corresponding standard trajectory response template, which includes the target control instruction sequence when an expert pilot performs standard maneuvers under the same weather and mission conditions. The system will collect the actual operation command sequence in real time. With target manipulation command sequence The input is fed into the phase difference analysis module, which performs second-order derivative operations on both sequences to identify instruction transition points where the rate of change of the instruction changes abruptly. The module then calculates the timestamps of these transition points in the actual sequences. Timestamps of corresponding transition points in the target sequence The difference between them generates the command response phase difference, which represents the perceived lead time. The calculation formula is as follows: ,in, The phase difference between the command and response, in seconds; This refers to the actual timing of the control command transition. When the target manipulation command jumps, A negative value indicates control lead, while a positive value indicates control lag; the system performs gaze-motion coherence determination, and extracts the gaze probability distribution time series from the video data of the instrument's gaze area. and motion trajectory functions in hand movement video data. The system defines the gaze switching frequency as the driving signal and the joystick displacement frequency as the response signal. It uses a cross-spectral density analysis method based on Fast Fourier Transform to calculate the frequency domain correlation between the two signals. The system calculates the correlation between the two signals in the preset focus frequency band, such as... to The amplitude squared coherence function within the frequency band is defined, and the mean of the coherence function within this frequency band is defined as the coherence coefficient. , The range of values ​​is to ,when Below the preset coupling threshold, such as When the manipulation frequency is higher than the standard template frequency, the system determines that there is a compensatory manipulation where the visual search logic and the manipulation execution action are separated.

[0031] The system calculates the cognitive load weighting coefficient, and the system responds to the command phase difference. With coherence coefficient Cognitive load weighting coefficients are generated using a weighted calculation model. The calculation formula is as follows: In the formula, This refers to the cognitive load weighting coefficient. This is the phase difference sensitivity parameter; For coherence sensitivity parameters, and Determined through offline calibration, i.e., by selecting expert pilot data. The mean approaches And select beginner error data to make Approaching The system generates a manipulation stability index, and the system will combine the feature vectors. Input to a pre-trained Transformer neural network model generates behavior representation vectors through nonlinear mapping. The system utilizes cognitive load weighting coefficients. Perform a weighted correction operation on the temporal variation trend of the behavior representation vector, and calculate the corrected change. The calculation formula is as follows: In the formula, This represents the change in the corrected behavior vector. This is the gain coefficient, used to adjust the magnitude of the weight's influence. and These are the behavior representation vectors for the current time step and the previous time step, respectively. The system calculates the manipulation stability index based on the corrected changes. The calculation formula is as follows: ,in, To manipulate the stability index; The Euclidean norm of a vector; As a normalization constant, this correction operation amplifies the variation in behavioral vectors when the cognitive load weighting coefficient is high, leading to a decrease in the manipulation stability index. Finally, based on the corrected behavioral representation vector sequence, the system uses a Long Short-Term Memory (LSTM) network or a Temporal Convolutional Network (TCN) to identify the trainee's manipulation style in the current task phase, classifying it as anticipatory, lagging, or over-controlling, and combining this with the manipulation stability index. Output the comprehensive evaluation results.

[0032] Example 1: In an instrument landing system approach and landing training scenario simulating strong crosswinds, the trainee maintains the aircraft within the preset glide slope and heading direction range through high-frequency correction operations. The flight trajectory deviation parameters meet the assessment pass standard, but the control stick displacement frequency is higher than the conventional benchmark. The system executes data synchronously, in order to... The global sampling frequency is used to time-align the flight fast access recorder parameters, hand movement video, and instrument gaze area video to generate a joint feature vector containing real-time flight attitude and control commands. The system initiates the phase-locked control entropy measurement procedure, loads the corresponding standard trajectory response template based on the current approach and landing mission semantics, including the target control command sequence for expert pilots to perform standard correction maneuvers under similar crosswind conditions. The system calculates the actual sequence of control commands. The time difference between the command response and the target sequence at the second derivative transition point is used to obtain the command response phase difference. In the calculation results for this working condition, It presents as a continuous sequence of positive values.

[0033] The system performs coherence measurements of gaze and movement, and extracts the temporal probability distribution of gaze points in the instrument's gaze region. With hand movement trajectory function Through cross-spectral density analysis, the system calculates the cross-spectral density of the two. to Coherence coefficient within the frequency band The calculation data shows the current time period Continuously below the preset coupling threshold This indicates that the trainees' eye-scanning did not effectively guide their hand movements, and that there was a decoupling between visual search logic and physical execution, confirming that the current high-frequency manipulation was a blind compensatory behavior. Based on the above quantitative indicators, the system used a weighted calculation model to derive a high cognitive load weight coefficient. In generating the manipulation stability index In the process, the system utilizes this behavior representation vector The rate of change is nonlinearly weighted and corrected according to the formula. higher Amplified and corrected change in behavior vector This leads to the final output handling stability index. The system reduces the speed and outputs an evaluation result, classifying the flight that met the trajectory as a delayed compensatory maneuver.

[0034] Example 2: This study constructed a high-fidelity flight training experimental platform to verify the effectiveness and stability of the proposed behavioral feature extraction and manipulation pattern analysis method in a real engineering environment. The core of the experimental platform is a full-motion flight simulator with an integrated sampling rate of [missing information]. High-precision eye-tracking system with positioning accuracy of The hand motion capture system can synchronously acquire the gaze coordinates and hand spatial displacement of the trainee with millisecond-level accuracy. Furthermore, the system uses the ARINC429 bus to... The frequency records QAR flight parameters, including pitch angle, roll angle, control stick displacement, and thrust stick position, in real time. To simulate complex disturbances in real flight, an unsteady sidewind model is actively introduced into the experimental environment, and a signal-to-noise ratio of [value missing] is superimposed in the sensor signal link. Gaussian white noise was used to test the algorithm's robustness under non-ideal conditions. The experimental design included two control groups: a control group and an experimental group based on the present invention. The control group consisted of 10 trainees in the initial training stage, representing maneuvering behavior without a mature perception model. The experimental group consisted of 10 skilled trainees who had completed advanced training, representing ideal maneuvering behavior with predictive capabilities. The experimental task was uniformly set as a precision approach and landing under strong crosswind conditions, with the crosswind intensity set at [insert value here]. This is a critical operating condition that places high demands on sensing lead time and handling stability.

[0035] Experiment to examine command response phase difference This core feature, the ability to recognize control patterns, was demonstrated by the system collecting the control command sequences of all subjects during the approach correction phase and comparing them with a standard track response template generated based on expert operation. Data analysis showed that the comparison sample group... The distribution exhibits a positive bias, with a mean of 1. This indicates that the control actions lag behind changes in flight status, exhibiting typical deviation-correction feedback characteristics. In contrast, the sample group of this invention... The mean stabilizes at to Furthermore, the experiment, through gaze-motor coherence analysis, delves into the underlying mechanism of compensatory manipulation and systematically calculates the temporal probability distribution of gaze regions. With joystick motion trajectory function exist to Cross-correlation coherence coefficient within the frequency band As shown in Table 1 below.

[0036] Table 1: Comparison of Characteristic Parameters of Control Modes

[0037]

[0038] See Table 1. Although the final trajectory deviation of the comparison sample group is... It barely meets the assessment criteria, but its coherence coefficient is only [missing information]. It is far below the preset coupling threshold, and the peak value of the manipulation frequency reaches This data set demonstrates that novice learners' high-frequency manipulation lacks effective visual guidance and constitutes blind compensatory correction. Conversely, the sample group of this invention maintains... While maintaining high trajectory accuracy, the coherence coefficient is also as high as [missing information]. And the manipulation frequency is controlled at .

[0039] Example 3: This example combines Figures 1 to 3 This paper describes a method for extracting behavioral features and analyzing control patterns for flight training. Figure 1 As shown, the system initiates multimodal data synchronous acquisition, which involves simultaneously acquiring QAR parameters, actions, and gaze videos and loading trajectory templates. It also performs parallel command response phase comparison calculations and visual-motor logic coherence measurements. The former generates a phase difference characterizing the trainee's perceived lead time through command jump point analysis, while the latter calculates the cross-correlation coherence coefficient between gaze switching and control frequency. The analysis results are input to the cognitive load weighting coefficient calculation module, which performs perceptual load quantification based on the phase difference and coherence coefficient. Simultaneously, the system uses behavioral embedding space mapping to transform multi-source feature vectors into a preset comprehensive behavioral representation vector, which then enters the behavioral representation vector weighting correction stage. Weighting coefficients are used to perform nonlinear weighting on the temporal trend, ultimately achieving a comprehensive control style determination after removing compensatory components and outputting a control stability index.

[0040] like Figure 2 As shown, the horizontal axis represents time in seconds, ranging from 0 to 9.6 seconds, and the vertical axis represents the normalized feature value ranging from 0 to 1.0. The graph uses solid lines to depict the fluctuation patterns of flight parameters, short dashed lines to reflect the displacement trends of hand movements, dotted lines to depict the temporal shifts in the gaze region, and long dashed lines to represent the feature changes of speech data. This visually presents the temporal alignment effect of multi-source features from different dimensions under a unified clock synchronization mechanism. Figure 3As shown, the complete system architecture consists of a multi-source sensing acquisition field, a full-clock synchronization gateway, an intelligent analysis core, and a comprehensive evaluation application field. The multi-source sensing acquisition field includes a flight parameter unit for recording the command sequence, a visual attention unit for the timing of gaze distribution, a limb movement unit for hand trajectory recognition, and a psychological stress monitoring unit for the extraction of speech acoustic indicators. All sensing data undergoes unified timing and data alignment preprocessing through the full-clock synchronization gateway before being input into the intelligent analysis core, which serves as the core logic engine. This core integrates four key functional modules: a command phase comparison engine, a visual-motor coherence analyzer, a cognitive load weight calculation, and a behavioral representation vector correction. The analysis results are ultimately projected onto the comprehensive evaluation application field, displaying the indicators after removing compensatory false components through a stability indicator dashboard. Furthermore, the system classifies trainees into predictive, lagging, or over-controlling operation styles through manipulation style profiling, and outputs evaluation confidence labels to provide multi-dimensional feature fusion verification support.

[0041] Example 4: This example focuses on the parameters used in calculating the cognitive load weighting coefficient. and To supplement the determination mechanism in the aforementioned weighted calculation model In the middle, the phase difference sensitivity parameter Coherence sensitivity parameter The value of determines the system's sensitivity to compensatory manipulation. This embodiment constructs an offline parameter calibration procedure based on the data envelopment analysis concept and establishes a benchmark database containing two types of typical samples. The first type is an expert benchmark set, consisting of no less than A senior flight instructor, performing a baseline approach under standard weather conditions, demonstrated precise and predictive maneuvers, theoretically possessing a cognitive load weighting coefficient. It should approach The second type is the typical compensation set, consisting of no less than The data from a group of novice trainees, exhibiting high-frequency corrections but barely meeting trajectory standards under strong interference conditions, is characterized by phase lag and look-move separation. Theoretically, its... It should approach The system extracts the absolute value of the command response phase difference of all samples in the benchmark database. With coherence coefficient The time series mean.

[0042] The system performs parameter optimization iterations, and the objective function is set as follows: The aim is to maximize the performance of both types of sample sets. Spatial separation, the objective function is defined as the degree of separation of the typical compensatory set. Mean and expert benchmark set The difference in means is used to introduce constraints and limit the concentration of expert benchmarks. Sample Value not exceeding To ensure a low false alarm rate, the system uses a genetic algorithm. and exist A search was conducted within the specified interval. During an actual calibration test for a specific type of fixed-wing trainer aircraft, the results were... After iterating through multiple generations of the population, the algorithm converges and outputs the optimal parameter combination. , Based on the calibration results, the system verified the stability of the parameters by applying them to a set of uncalibrated test data, including the operation records of trainees with mixed skill levels. The results showed that the system was able to achieve [the desired results]. The accuracy rate correctly distinguishes between predictive and compensatory maneuvers, and the consistency coefficient of the assessment under different crosswind intensities reaches [percentage missing]. .

[0043] Example 5: To ensure the universality and stability of the cognitive load weighting coefficient calculation model, the phase difference sensitivity parameter needs to be adjusted. Coherence sensitivity parameter Following standardized initial calibration procedures, the system constructs a baseline database containing two types of typical control samples: one type consists of data from expert pilots performing baseline approaches under standard weather conditions, characterized by precise and predictive control; the other type consists of data from novice pilots maintaining trajectory compliance solely through high-frequency corrections under strong crosswind interference, characterized by phase lag and visual-motor separation. The system calculates the absolute value of the phase difference between command responses for these two types of samples. With coherence coefficient The statistical distribution characteristics of the data.

[0044] Based on the benchmark database, the system uses an offline optimization algorithm to determine... and The optimal value is set by maximizing the calculated cognitive load weight coefficients of the two classes of samples. The discriminatory power, that is, not only the expert sample The mean approaches At the same time, it is necessary to make the compensatory sample The mean approaches Before actual deployment, the system iterates within a preset parameter space using a genetic algorithm or grid search until it finds the parameter combination that maximizes the discrimination index. This procedure ensures that the system can adaptively adjust the sensitivity of the evaluation model according to the characteristics of different models or training subjects, thereby eliminating evaluation bias caused by improper parameter settings.

[0045] Example 6: This example provides supplementary explanation of the engineering determination method for nonlinear correction parameters in the generation of the handling stability index. In the specific implementation, the handling stability index... The calculation depends on the amount of change in the modified behavior vector. In the formula, the gain coefficient Determine the cognitive load weighting coefficient To mitigate evaluation bias caused by improper parameter settings, this embodiment discloses a gain coefficient based on the principle of balancing false alarm rate and false negative rate to address the penalty for the stability index. On-site calibration procedures; executed during the initial system deployment phase, baseline calibration is performed using a pre-defined standard test case set. This test set contains two sets of labeled data: one set consists of high-stability, low-compensation samples, i.e., normal operation data of expert pilots under minor disturbances; the other set consists of low-stability, high-compensation samples, i.e., overcorrection data of novice trainees under strong disturbances. The system... Set the distance from the walk within the range, and iterate through and calculate each group. The stability index distribution under each value, for each The system calculates the area of ​​the overlapping region between the two types of samples and uses minimizing this overlapping area as the optimization objective.

[0046] Determine the initial selection After the initial value is set, the system enters the online adaptive fine-tuning phase. During actual training, the system monitors the consistency between the evaluation results and the instructor's subjective scores in real time. When continuous consistency is detected... There is a discrepancy between the evaluation results of each training session and the instructor's rating, i.e., the correlation coefficient is lower than [missing value]. At that time, the system automatically triggers parameter fine-tuning logic, adjusting the parameters according to the direction of deviation. The value is slightly adjusted, with a step size not exceeding [missing value]. This continues until the consistency index returns to above the preset threshold.

[0047] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for extracting behavioral features and analyzing control patterns for flight training, characterized in that, Includes the following steps: Step S1: Simultaneously collect flight quick access recorder parameters, hand movement video data, and instrument gaze area video data during flight training, and load the corresponding standard trajectory response template according to the flight mission stage. Step S2: Analyze the control command sequence in the flight quick access recorder parameters, and perform phase comparison calculation between the control command sequence and the standard track response template to generate the command response phase difference that characterizes the trainee's control perception lead. Step S3: Use the action recognition model to extract the motion trajectory function from the hand action video data, and use the visual attention model to determine the temporal probability distribution of the gaze point in the instrument gaze area video data. Step S4: Calculate the cross-correlation degree between the temporal probability distribution of the gaze point and the spectral characteristics of the motion trajectory function, and determine the coherence coefficient representing the degree of coupling between the visual search logic and the manipulation execution action based on the cross-correlation degree; Step S5: Calculate the cognitive load weighting coefficient based on the command response phase difference and coherence coefficient, and convert the flight rapid access recorder parameters and video data into a preset behavior embedding space through a nonlinear mapping algorithm to generate a behavior representation vector that represents the trainee's comprehensive operational state. Step S6: Perform a weighted correction operation on the temporal change trend of the behavior representation vector using the cognitive load weight coefficient to generate the manipulation stability index after removing the compensatory manipulation component.

2. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, When generating the command response phase difference, the following operations are performed: second derivative operations are performed on the target control command sequence contained in the standard track response template and the acquired actual control command sequence to identify command jump points where the command rate of change changes abruptly; characteristic timestamps of the actual control command sequence at the command jump points are extracted, and the time offset of the characteristic timestamp relative to the corresponding jump point in the target control command sequence is calculated; the average offset value of multiple command jump points is used to determine the trainee's perception lead in the current flight mission phase.

3. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, Step S4 specifically includes: defining the gaze switching frequency in the probability distribution time series as the driving signal, defining the manipulation frequency corresponding to the motion trajectory function as the response signal; determining whether the manipulation behavior is controlled by the visual search logic by analyzing the overlap of the energy distribution of the driving signal and the response signal in the frequency domain, and defining the value of the overlap of the energy distribution as the coherence coefficient.

4. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, When calculating the cognitive load weighting coefficient in step S5, the following quantification rules are followed: ,in, For cognitive load weighting coefficient, For the command response phase difference, The coherence coefficient, as well as These are the preset sensitivity adjustment parameters.

5. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, Step S6 specifically includes: real-time monitoring of the numerical change of the coherence coefficient; if the coherence coefficient is lower than the preset coupling threshold and the command response phase difference shows a non-linear fluctuation trend, then it is determined that the trainee's current stage of manipulation behavior is blind compensatory manipulation.

6. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 5, characterized in that, In cases where the manipulation is determined to be blind compensatory, the rate of change of the behavioral representation vector is doubled using the cognitive load weighting coefficient to reduce the spurious stability weights generated by blind compensation in the manipulation stability index.

7. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, After generating the control stability index, the following operations are also included: acquiring voice audio data during flight training and extracting acoustic indicators that characterize psychological stress from the voice audio data; using the acoustic indicators to calculate the confidence level of the control stability index and attaching a corresponding confidence level label to the control stability index.

8. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 7, characterized in that, The steps for extracting acoustic indicators include: analyzing the pitch change rate, pause length, and fundamental frequency fluctuation amplitude of the speech audio data; if any parameter of the pitch change rate, pause length, or fundamental frequency fluctuation amplitude exceeds the preset stress threshold, the weight ratio of the cognitive load weight coefficient in the correction calculation is increased.

9. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, Also includes: We analyzed the long-sequence characteristics and trends of the manipulation stability index during the training process and classified the manipulation behavior of trainees into predictive, lagging, or over-controlling manipulation styles.

10. The method for extracting behavioral features and analyzing control patterns for flight training according to claim 1, characterized in that, In step S1, the collected flight quick access recorder parameters, hand movement video data, instrument gaze area video data, and voice audio data are time-aligned through a unified clock synchronization mechanism so that the manipulation behavior at any time corresponds to a unique joint feature vector.

Images