Deep learning based method for evaluating flight operation quality of an aviation simulator

Abstract

The application relates to the technical field of management evaluation, and discloses an aviation simulator flight operation quality evaluation method based on deep learning, which comprises the following steps: acquiring control input data, attitude response data and condition metadata, calculating transmission delay and constructing device condition parameters; performing delay mapping correction on the attitude data based on the transmission delay, calculating control pulse degree and response sensitivity, and then splicing and constructing comprehensive multi-dimensional time sequence data; inputting the time sequence data into a deep space-time information extraction network, and adopting a particle swarm algorithm to globally optimize hyperparameters; in the optimization process, a particle swarm dispersion coefficient is calculated by extracting an individual optimal position set, continuous modulation is performed on inertia weight and divergence cost weight, and optimization distribution dispersion induced by the device condition parameters is inhibited; and finally, network training is completed based on the optimal hyperparameters, and a sliding window reasoning output rating label is output for an examination sequence. The application eliminates the interference of cross hardware configuration differences and improves the objective consistency of capability evaluation.

Description

Technical Field

[0001] This invention relates to the field of management evaluation technology, and more specifically, to a deep learning-based method for evaluating the quality of flight operations in an aviation simulator. Background Technology

[0002] In modern multi-level flight simulator training systems, competency-based and evidence-based training standards are becoming increasingly prevalent. Factors such as instructor support levels, equipment levels, and situational complexity have been explicitly incorporated into the evaluation dimensions of pilot operational quality and are closely linked to training difficulty in actual scheduling. In recent years, using deep learning to model and automatically rate pilot maneuver data has become an important development direction in this field. Existing deep learning evaluation methods typically directly input the collected maneuver inputs and response sequences into a neural network and rely on online algorithms (such as particle swarm optimization) to optimize and train the network's hyperparameters. However, flight simulators are closed-loop human-machine systems subject to strict regulatory constraints. The underlying hardware configurations of different levels of equipment have inherent objective differences, mainly reflected in end-to-end transmission delays, dynamic damping feedback of the control loading system, and phase lag between the motion platform and the visual display. These systemic differences in hardware architecture can lead to drastically different micro-manipulation compensation behaviors from pilots on different devices, resulting in a severe stratification phenomenon in the generation mechanism of training data collected from multiple devices. Existing evaluation methods often treat these objective hardware differences as ordinary noise when processing this multi-source data, failing to achieve effective isolation. When using particle swarm optimization (PSO) to optimize hyperparameters in complex multi-peak environments, the optimization terrain induced by different hardware parameters is prone to divergence and drift. This phenomenon forces deep learning networks to follow the underlying hardware architecture of the simulator to extract distribution information during training, leading to a technical defect where the rating decision function for the same capability objective shifts significantly with changes in hardware conditions. Summary of the Invention

[0003] This invention provides a deep learning-based method for evaluating the flight operation quality of an aviation simulator, which solves the technical problems mentioned in the background.

[0004] This invention provides a deep learning-based method for evaluating the flight operation quality of an aviation simulator, including:

[0005] Acquire the flight simulator's control input data, attitude response data, simulator condition metadata, and assessment sequences;

[0006] Calculate the transmission delay between the manipulation input data and the attitude response data, and construct device condition parameters based on the transmission delay and the simulator condition metadata;

[0007] Delay mapping correction is performed on the attitude response data based on the transmission delay to obtain corrected attitude data; control impulse degree and response sensitivity are calculated based on the manipulation input data and the corrected attitude data; the manipulation input data, the corrected attitude data, the control impulse degree, the response sensitivity, and the device condition parameters are concatenated to construct comprehensive multidimensional time series data;

[0008] The comprehensive multidimensional time series data is input into the deep spatiotemporal information extraction network, and the hyperparameters of the deep spatiotemporal information extraction network are globally optimized using the particle swarm optimization algorithm.

[0009] During the global optimization process, the set of individual optimal positions of the particle swarm optimization algorithm is extracted to calculate the particle swarm discrete coefficients.

[0010] The inertia weight of the particle swarm algorithm and the divergence cost weight in the preset optimization objective function are continuously modulated based on the particle swarm discrete coefficients to suppress the optimization distribution discrepancy of the hyperparameters induced by the device condition parameters.

[0011] The deep spatiotemporal information extraction network is trained based on the optimal hyperparameters obtained after modulation convergence, so as to perform sliding window inference on the assessment sequence and output rating labels.

[0012] The beneficial effects of this invention are as follows: by acquiring device condition parameters, including transmission delay, to perform delay mapping correction on attitude data, and by combining control impulse degree and response sensitivity to construct comprehensive multi-dimensional time-series data, the data alignment error caused by differences in the underlying hardware configuration of multi-level simulators and transmission link delay is effectively eliminated. At the same time, in the process of using particle swarm optimization to globally optimize the hyperparameters of deep networks, the particle swarm discrete coefficients of the individual optimal position set are extracted, and the inertia weight and the divergence cost weight in the optimization objective function are continuously modulated accordingly. This suppresses the hyperparameter optimization distribution discreteness problem induced by the objective hardware parameters of the simulator, so that the evaluation model finally trained can completely remove the evaluation noise caused by the coupling of equipment system changes and scheduling difficulty. This achieves high-precision, high-robustness, and objective and consistent rating output of flight trainees' real flying ability across devices and training stages. Attached Figure Description

[0013] Figure 1 This is a flowchart of the flight operation quality evaluation method for aviation simulators based on deep learning, as described in this invention.

[0014] Figure 2 This is a graph showing the experimental data of the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0016] like Figure 1 As shown, a deep learning-based method for evaluating the flight operation quality of an aviation simulator includes:

[0017] Acquire the flight simulator's control input data, attitude response data, simulator condition metadata, and assessment sequences;

[0018] Calculate the transmission delay between the manipulation input data and the attitude response data, and construct device condition parameters based on the transmission delay and the simulator condition metadata;

[0019] Delay mapping correction is performed on the attitude response data based on the transmission delay to obtain corrected attitude data; control impulse degree and response sensitivity are calculated based on the manipulation input data and the corrected attitude data; the manipulation input data, the corrected attitude data, the control impulse degree, the response sensitivity, and the device condition parameters are concatenated to construct comprehensive multidimensional time series data;

[0020] The comprehensive multidimensional time series data is input into the deep spatiotemporal information extraction network, and the hyperparameters of the deep spatiotemporal information extraction network are globally optimized using the particle swarm optimization algorithm.

[0021] During the global optimization process, the set of individual optimal positions of the particle swarm optimization algorithm is extracted to calculate the particle swarm discrete coefficients.

[0022] The inertia weight of the particle swarm algorithm and the divergence cost weight in the preset optimization objective function are continuously modulated based on the particle swarm discrete coefficients to suppress the optimization distribution discrepancy of the hyperparameters induced by the device condition parameters.

[0023] The deep spatiotemporal information extraction network is trained based on the optimal hyperparameters obtained after modulation convergence, so as to perform sliding window inference on the assessment sequence and output rating labels.

[0024] Preferably, the acquisition includes the flight simulator's control input data, attitude response data, simulator condition metadata, and assessment sequences, including:

[0025] Get timestamps with control input Lateral displacement of the control stick Longitudinal displacement of the control stick Pedal deflection and throttle status Construct the manipulation input data Its expression is:

[0026]

[0027] Get the timestamp of the instrument display refresh Visual frame timestamp and motion command output timestamp attitude angle, angular velocity, airspeed and height Construct the attitude response data Its expression is:

[0028]

[0029] in, , , Indicates the attitude angle, , , This indicates the angular velocity;

[0030] Obtaining Instructor Support Level Context complexity level Equipment identification Student identification and subject identification , as the simulator condition metadata;

[0031] Obtain the timing data of the manipulation and attitude to be evaluated, and use it as the assessment sequence. .

[0032] The control input timestamp is a record of the time when the flight simulator receives the pilot's control input signals. It can be obtained by synchronizing with the system's high-precision real-time clock through the flight simulator's hardware interface acquisition module.

[0033] Lateral displacement of the control stick is the amount of displacement generated when the pilot performs lateral operation on the control stick of the flight simulator. It can be collected by contact or non-contact displacement sensors mounted on the control stick of the flight simulator.

[0034] The longitudinal displacement of the control stick is the amount of displacement generated by the pilot's longitudinal operation of the control stick in the flight simulator, which can be collected by the contact or non-contact displacement sensor mounted on the flight simulator control stick.

[0035] Pedal deflection is the deflection angle generated by the pilot operating the pedal on the flight simulator, which can be collected by the angle sensor mounted on the pedal on the flight simulator.

[0036] Throttle status is the positional state of the pilot's operation of the throttle control device on the flight simulator, which can be collected by the position sensor mounted on the throttle device of the flight simulator.

[0037] The instrument display refresh timestamp is a record of the time when the flight simulator instrument module completes the data refresh. It can be obtained by synchronizing the refresh control module of the flight simulator instrument display system with the system clock.

[0038] The visual frame timestamp is a time record of when the flight simulator vision system outputs a visual image frame. It can be obtained by synchronizing the frame output control unit of the flight simulator image generation module with the system clock.

[0039] The motion command output timestamp is a record of the time when the motion system of the flight simulator receives and outputs motion control commands. It can be obtained by synchronizing the command sending unit of the motion control module of the flight simulator with the system clock.

[0040] Attitude angles are the collective term for the pitch, roll, and yaw angles of an aircraft in space, as simulated by an aviation simulator. They can be collected and output in real time through the flight simulation calculation module of the aviation simulator.

[0041] Angular velocity is a general term for the rotational angular velocity of an aircraft around its three axes as simulated by an aviation simulator. It can be collected and output in real time through the flight simulation calculation module of the aviation simulator.

[0042] Airspeed is the speed of an aircraft relative to the air as simulated by an aviation simulator, and it can be collected and output in real time through the flight simulation calculation module of the aviation simulator.

[0043] Altitude is the vertical distance between the simulated aircraft and a specified reference plane in the flight simulator, and it can be collected and output in real time through the flight simulation solution module of the flight simulator.

[0044] Instructor support level is a classification of the level of operational support provided by instructors to trainees during flight simulator training. It can be manually entered and collected through the instructor operation station system of the flight simulator.

[0045] The situation complexity level is a classification of the complexity of the environment and tasks set in the training subjects of the flight simulator. It can be pre-configured and collected through the training management system of the flight simulator.

[0046] Equipment identifiers are unique identifiers used to distinguish different flight simulator equipment, and can be pre-assigned and collected through the flight simulator's equipment management system.

[0047] Trainee identification is a unique identifier used to distinguish different trainees, and it can be collected through the trainee information module of the training management system of the flight simulator.

[0048] Subject identification is a unique identifier used to distinguish different flight training subjects, which can be collected through the subject configuration module of the training management system of the flight simulator.

[0049] The assessment sequence is a collection of timing data of control inputs and attitude responses collected by the flight simulator during the student assessment phase. It can be collected and stored separately in assessment mode by the flight simulator's data acquisition system.

[0050] In detail, the process involves binding the manipulation input quantities with the control input timestamps during collection to match the timing of subsequent attitude responses, avoiding timing misalignments caused by multi-link transmission delays. The aircraft simulator's hardware interface associates and stores the timestamps of each manipulation input sampling point with the system's high-precision clock in real time, forming a combination of physical quantity and timestamp data for each sampling point. Simultaneously, the attitude response physical quantities are bound to the timestamps of the instrument display refresh, visual frames, and motion command output channels, adapting to the multi-module serial structure of the aircraft simulator. Each channel's device module generates a synchronization timestamp upon completing data refresh or command output, ensuring that the response data of each channel is traceable to the corresponding manipulation input time. Furthermore, training organization condition metadata such as instructor support level and situation complexity level are collected collaboratively with device, student, and subject identifiers. All metadata is associated with the same training session identifier for the corresponding manipulation and attitude timing data, ensuring a one-to-one correspondence between training conditions and operational data. Finally, the manipulation and attitude timing data to be evaluated are separately divided into assessment sequences, which are marked with unique assessment identifiers and stored separately from daily training data.

[0051] In detail, displacement of the control inputs is measured in millimeters, angles in degrees, airspeed in meters per second, and altitude in meters. All physical quantities are normalized according to the industry standard for aviation simulators. The time stamp acquisition accuracy is at the 1-millisecond level, relying on the high-precision real-time clock of the aviation simulator system. The time stamp generation units of each module are hard-synchronized with this clock. The acquisition frequency of control inputs and attitude response data is 100 Hz, ensuring that the acquired timing data can fully reflect the micro-control characteristics of the pilot. Instructor support levels are divided into 0 to 3 levels. Level 0 is no prompts, no takeover, and only safety boundaries are maintained. Level 1 is a few prompts, with no more than 3 prompts and no takeover. Level 2 is multiple prompts, with more than 3 prompts and possible short-term takeover, with no more than 2 takeover attempts. Level 3 is high support. The instructor provides continuous guidance or frequent takeover, and this classification is entered by the instructor through the instructor operation station before training. The situation complexity level is divided into 1 to 5 levels: Level 1 is basic maneuver subjects without complex environment; Level 2 is basic maneuver subjects plus simple weather; Level 3 is instrument approach subjects without faults; Level 4 is instrument approach subjects plus complex weather; and Level 5 is complex weather plus high workload subjects with fault injection. This classification is pre-configured by the training management system according to subject type. Equipment identification adopts a unique digital coding rule, with each flight simulator assigned a unique positive integer code. The coding information is stored in the equipment management system and bound to the equipment hardware. The duration of the assessment sequence is no less than 60 seconds, and the corresponding data volume is no less than 6,000 sampling points to ensure that the collected assessment data has sufficient operational characteristics for flight operation quality evaluation.

[0052] Preferably, the process of calculating the transmission delay between the manipulation input data and the attitude response data, and constructing device condition parameters based on the transmission delay and the simulator condition metadata, includes:

[0053] Calculate the refresh timestamp of the instrument display respectively. The visual frame timestamp and the motion command output timestamp With the control input timestamp The difference is used to obtain the instrument channel delay sequence. Visual channel delay sequence and motion channel delay sequence Its expression is:

[0054]

[0055]

[0056]

[0057] The median of the instrument channel delay sequence, the visual channel delay sequence, and the motion channel delay sequence are extracted respectively, and used as the robust transmission delay of the instrument. Visually robust transmission delay and motion-steady transmission latency Its expression is:

[0058]

[0059]

[0060]

[0061] in, This represents the median extraction operation along the time dimension; if the motion command output timestamp does not exist, then the visual robust transmission delay is used as the motion robust transmission delay, i.e. ;

[0062] Based on the total number of devices Identification of the device Normalization is performed to obtain the device code identifier. Its expression is:

[0063]

[0064] in, This represents the integer number corresponding to the device identifier;

[0065] The instrument robust transmission delay, the visual robust transmission delay, the motion robust transmission delay, and the instructor support level are all considered. The aforementioned situation complexity level The device condition parameters are constructed by combining the device code identifier with the device code identifier. Its expression is:

[0066]

[0067] The instrument channel delay sequence is a time-series data composed of the time difference between the instrument display refresh timestamp and the control input timestamp on the flight simulator.

[0068] The visual channel delay sequence is a time-series data composed of the time difference between the visual frame timestamp and the control input timestamp of the flight simulator.

[0069] The motion channel delay sequence is a time-series data composed of the time difference between the motion command output timestamp and the control input timestamp of the flight simulator.

[0070] The robust transmission delay of an instrument is the median extracted from the delay sequence of the instrument channel, and is used to characterize the robustness of the transmission delay of the instrument channel.

[0071] Visual robust propagation delay is the median extracted from the visual channel delay sequence and is used to characterize the robustness of visual channel propagation delay.

[0072] Motion robust transport delay is the median extracted from the motion channel delay sequence and is used to characterize the robustness of motion channel transport delay. When there is no motion command output timestamp, the visual robust transport delay is used.

[0073] The total number of devices refers to the total number of aviation simulator devices participating in training or assessment, which can be collected and statistically analyzed through the aviation simulator's device management system.

[0074] The device identification integer number is a unique integer code converted from the device identification, used for the digital differentiation of devices.

[0075] The equipment code identifier is a standardized code obtained by normalizing the integer number of the equipment identifier according to the total number of equipment, and the value range is between 0 and 1.

[0076] The device condition parameters are multi-dimensional comprehensive parameters that integrate instrument robust transmission delay, vision robust transmission delay, motion robust transmission delay, instructor support level, situation complexity level, and device coding identifier. They are used to quantify the structural characteristics and training conditions of the simulator.

[0077] In detail, calculating the timestamp difference across channels to obtain the latency sequence for each link is crucial because the instrument, vision, and motion systems of an aircraft simulator are independent modules with different transmission delay characteristics. Channel-specific calculation accurately captures the structural differences between these modules. For example, the instrument channel latency is typically 40 to 80 milliseconds, the vision channel latency is 60 to 100 milliseconds, and the motion channel latency is 80 to 150 milliseconds. Channel-specific recording preserves the specificity of these modules. The median is chosen as the robust transmission latency because instantaneous latency jitter may occur during simulator operation. The median effectively suppresses extreme value interference and reflects the true latency level better than the mean. For example, in a training period, the instrument channel latency sequence includes 30 milliseconds, 45 milliseconds, 50 milliseconds, 100 milliseconds, and 48 milliseconds. The median of 48 milliseconds more closely reflects the actual transmission latency than the mean of 54.6 milliseconds. Using the robust vision transmission latency instead when there is no motion command output timestamp is a fixed rule designed for low-level simulators without a motion module, ensuring compatibility across different device configurations. The device parameters are standardized. For example, the FNPT level simulator has no motion system, so its visual robust transmission latency of 75 milliseconds is directly used as the motion robust transmission latency. Device identifiers are normalized to eliminate inconsistencies in encoding scale caused by differences in the total number of devices. For example, the identifier integers of 3 devices (1, 2, 3) are normalized to 0, 0.5, and 1, respectively. The identifier integers of 5 devices (1 to 5) are normalized to 0, 0.25, 0.5, 0.75, and 1, ensuring comparability of encodings across different device scales. Device condition parameters are constructed by combining multiple dimensions because these parameters collectively determine the generation mechanism of training data. Integrating them into a unified set of parameters allows subsequent deep networks to simultaneously perceive the simulator structure and training organization conditions. For example, the device condition parameters for a certain training scenario include instrument robust transmission latency of 45 milliseconds, visual robust transmission latency of 70 milliseconds, motion robust transmission latency of 70 milliseconds, instructor support level 1, situational complexity level 4, and device encoding identifier of 0.5, comprehensively quantifying the core conditions of the scenario.

[0078] In detail, the median extraction time window is the entire duration of a single training session, meaning all latency sampling points from the start to the end of training participate in the median calculation to ensure coverage of latency characteristics throughout the entire training process. For example, in a 60-minute training session, 360,000 instrument channel latency sampling points are collected, and the median is extracted from this sequence. The rule for converting device identifiers to integer numbers is to assign consecutive positive integers according to the registration order of the devices in the management system. The registration order is pre-set by the flight school or training center. For example, the first registered simulator is numbered 1, the second is numbered 2, and so on, ensuring a unique correspondence between the identifier and the integer number. All dimensions of device condition parameters need to be standardized in terms of units, and latency parameters (milliseconds) are divided by 10. 00 is converted to seconds. Instructor support level (0 to 3) and situation complexity level (1 to 5) are normalized to the 0 to 1 range by dividing by 3 and 5 respectively, to maintain the same dimension as the device code identifier (0 to 1). For example, the instrument robust transmission delay of 45 milliseconds is converted to 0.045, instructor support level 1 is converted to 0.333, and situation complexity level 4 is converted to 0.8. The timestamp matching rule adopts one-to-one sampling point matching, that is, the control input timestamp corresponding to each control input sampling point is matched one-to-one with the timestamp corresponding to the instrument display refresh, visual frame output, and motion command output triggered by that sampling point to calculate the difference, to ensure the temporal consistency of the delay sequence and avoid delay calculation errors caused by sampling asynchrony.

[0079] Preferably, delay mapping correction is performed on the attitude response data according to the transmission delay to obtain corrected attitude data; control impulse degree and response sensitivity are calculated based on the manipulation input data and the corrected attitude data; the manipulation input data, the corrected attitude data, the control impulse degree, the response sensitivity, and the device condition parameters are concatenated to construct comprehensive multi-dimensional time-series data, including:

[0080] Using a sampling step size corresponding to a uniform sampling frequency Interpolate the manipulation input data and the attitude response data to obtain equally spaced manipulation input data. With equally spaced attitude response data Based on the robust transmission delay of the instrument With the sampling step size The ratio of the two values ​​is used to offset the equally spaced attitude response data along the time axis to obtain the corrected attitude data. Its expression is:

[0081]

[0082] in, This indicates the rounding operation;

[0083] The equally spaced manipulation input data and the corrected attitude data are filtered and normalized respectively to obtain normalized manipulation input data. With normalized corrected attitude data ;

[0084] The normalized manipulation input data is differentially processed to obtain the derivative of the manipulation input data. ;

[0085] The ratio of the sum of the first norm of the derivatives of the manipulated input data to the sum of the second norm of the normalized manipulated input data is used to obtain the control impulse degree. Its expression is:

[0086]

[0087] in, Indicates the total number of sampling points. Represents the zero constant. This represents the sum of absolute values ​​with respect to a norm. Represents the L2 norm;

[0088] Calculate the differential change of the normalized corrected attitude data. The difference in the normalized manipulated input data The ratio of the values ​​is then smoothed to obtain the response sensitivity. Its expression is:

[0089]

[0090]

[0091] in, Indicates the original sensitivity ratio. Represents the exponential smoothing coefficient;

[0092] For the equipment condition parameters Perform linear mapping and tiling to obtain tiling condition parameters. Its expression is:

[0093]

[0094]

[0095] in, and These represent the linear mapping weight matrix and the linear mapping bias vector, respectively.

[0096] At each sampling time, the normalized manipulation input data, the derivative of the manipulation input data, the normalized corrected attitude data, the response sensitivity, the control impulse degree, and the tiling conditional parameters are concatenated to construct the comprehensive multidimensional time-series data. Its expression is:

[0097]

[0098] in, This indicates a splicing operation.

[0099] A uniform sampling frequency is a fixed sampling frequency used when interpolating control input data and attitude response data. It is preferably 100 Hz. This frequency can fully preserve the pilot's micro-control characteristics and simulator response dynamics, while taking into account both data volume and computational efficiency.

[0100] The sampling step size is the time interval corresponding to the uniform sampling frequency, which is calculated from the reciprocal of the uniform sampling frequency.

[0101] Equal-interval manipulation input data is time-series data with equal time intervals obtained by interpolating the original manipulation input data at a uniform sampling frequency.

[0102] Equal-interval attitude response data is time-series data with equal time intervals obtained by interpolating the original attitude response data at a uniform sampling frequency.

[0103] The corrected attitude data is time-series data obtained by offsetting the attitude response data on the same ground axis based on the ratio of instrument robust transmission delay to sampling step size. It is used to eliminate the control-response timing misalignment caused by transmission delay.

[0104] Normalized manipulation input data is standardized data obtained by filtering and normalizing equally spaced manipulation input data.

[0105] Normalized attitude correction data is standardized data obtained by filtering and normalizing the corrected attitude data.

[0106] The derivative of the manipulation input data is a rate-type data obtained by differentiating the normalized manipulation input data, and is used to characterize the rate of change of the manipulation input.

[0107] The total number of sampling points is the total number of equally spaced data points collected during a single training or assessment process, which is obtained by multiplying the sampling duration by the uniform sampling frequency.

[0108] The zero-prevention constant is a very small positive number used to avoid zero denominators during calculations. It is preferably 10 to the power of negative 6. This value is small enough that it will not affect the accuracy of the calculation results, and at the same time, it can effectively avoid the singular case of zero denominators.

[0109] Control impulse degree is the ratio of the sum of the first norm of the derivatives of the manipulation input data to the sum of the second norm of the normalized manipulation input data. It is used to quantify the micro-manipulation morphological characteristics induced by the simulator control loading system.

[0110] The normalized corrected attitude data difference change is the difference between the normalized corrected attitude data of adjacent sampling points, used to characterize the instantaneous change in attitude response.

[0111] The normalized manipulation input data difference change is the difference between the normalized manipulation input data of adjacent sampling points, used to characterize the instantaneous change of the manipulation input.

[0112] The raw response sensitivity is the ratio of the differential change in the normalized corrected attitude data to the differential change in the normalized control input data, used to initially quantify the correlation strength between the control input and the attitude response.

[0113] The exponential smoothing coefficient is a weighting coefficient used when smoothing the original response sensitivity. It is preferably 0.2. This value can effectively suppress instantaneous spike interference while preserving the temporal variation characteristics of the sensitivity.

[0114] Smoothed response sensitivity is time-series data obtained by exponentially smoothing the original response sensitivity, and is used to stably characterize the strength of the manipulation-response correlation.

[0115] The linear mapping weight matrix is ​​the weight matrix used when linearly mapping device condition parameters. It is preferably a randomly initialized 6-row, 16-column matrix because the device condition parameters are 6-dimensional. After linear mapping, they need to be adapted to the input dimension of the time series network. 16-dimensional is the preferred dimension in engineering practice that balances representation ability and computational cost.

[0116] The linear mapping bias vector is a bias vector used when linearly mapping device condition parameters. It is preferably a randomly initialized 16-dimensional vector that is consistent with the linear mapping weight matrix and is used to compensate for the offset of the linear mapping.

[0117] Embedded conditional parameters are standardized vectors obtained by linearly mapping device conditional parameters, used to adapt to the input requirements of time-series networks.

[0118] Tiling conditional parameters are time-series data obtained by copying the embedded conditional parameters to each sampling time, so that the static conditional parameters can be continuously perceived by the time-series network.

[0119] Multidimensional time-series data is a high-dimensional time-series data obtained by splicing normalized manipulation input data, manipulation input data derivatives, normalized corrected attitude data, smoothed response sensitivity, control impulse degree, and tiling condition parameters in a fixed order, and is used as input for a deep spatiotemporal information extraction network.

[0120] In detail, the attitude response data is offset along the time axis based on the ratio of the instrument robust transmission delay to the sampling step size. By quantizing the sampling point offset corresponding to the delay, precise timing alignment between the control input and the attitude response is achieved. For example, if the instrument robust transmission delay is 45 milliseconds and the sampling step size is 10 milliseconds, the offset corresponds to 4.5 sampling points, which is rounded to 5 sampling points, ensuring that each control input matches the attitude response it triggers. The calculation method for the control impulse degree is specifically designed for the force-displacement dynamic design of the control loading system of the flight simulator. The derivative norm 1 highlights the instantaneous changes in the control, while the position norm 2 reflects the overall amplitude of the control. The ratio can effectively distinguish the control habits under different control loading stiffnesses. For example, a high-stiffness system will result in a larger control derivative norm 1 and a higher control impulse degree. The response sensitivity is smoothed by adding an exponential smoothing to the ratio of differential changes, which preserves the... The temporal dynamics of the manipulation-response relationship suppress noise interference and adapt to the phase lag effects brought by simulator motion cues algorithms and vision systems. The linear mapping and tiling of equipment condition parameters first solves the problem of dimensionality adaptation between static condition parameters and dynamic temporal data, enabling the deep network to simultaneously perceive equipment structural features and training conditions. By splicing multiple types of data in a fixed order, the constructed comprehensive multidimensional temporal data integrates original manipulation features, rate features, response features, structure-induced features, and conditional features, ensuring that the deep network can comprehensively learn information related to flight operation quality. For example, the comprehensive multidimensional temporal data at a certain sampling moment includes 4-dimensional normalized manipulation input, 4-dimensional manipulation derivative, 8-dimensional normalized corrected attitude, 1-dimensional smoothed response sensitivity, 1-dimensional control impulse degree, and 16-dimensional tiling conditional parameters, totaling 34 features, which can completely characterize the operational state and environmental conditions at that moment.

[0121] In detail, the sampling frequency is fixed at 100 Hz, with a corresponding sampling step size of 0.01 seconds, ensuring that data collected by all devices is consistent with the same time base. A second-order Butterworth low-pass filter is used for filtering, with a cutoff frequency of 6 Hz for the manipulation input data and 4 Hz for the attitude correction data. Filtering is performed after interpolation to obtain equally spaced data and before normalization. This parameter combination effectively filters out high-frequency noise while preserving core operational and response characteristics. Normalization uses a robust normalization algorithm, calculating the median and absolute median difference of the data for a single training session. The median is subtracted from the data, and then divided by the sum of the absolute median difference and the zero-prevention constant to avoid the influence of outliers on the data distribution. For example, if the median of a manipulation input data sequence is 50 mm and the absolute median difference is 10 mm, the normalized data range will remain stable within a reasonable range. The derivative of the manipulation input data is calculated using the central difference method, i.e., the data of the next adjacent sampling point is subtracted from the... The data from the previous sampling point is divided by twice the sampling step size. The first and last sampling points of the sequence are subjected to forward and backward differencing, respectively, to ensure the accuracy and completeness of the derivative calculation. The difference step size of the difference change is fixed at one sampling interval, i.e., the difference between two adjacent sampling points. The linear mapping weight matrix and bias vector adopt the Hessian initialization method to ensure that the initial values ​​can make the network training stable and convergent, and are updated together with the network parameters during training. The splicing order of the comprehensive multidimensional time series data is fixed as normalized manipulation input data, manipulation input data derivative, normalized corrected attitude data, smoothed response sensitivity, control impulse degree, and tiling conditional parameters. The dimensions of each part are 4-dimensional, 4-dimensional, 8-dimensional, 1-dimensional, 1-dimensional, and 16-dimensional, respectively, with a total dimension of 34 dimensions, to ensure that the input data structure is consistent each time. The zero constant is fixed at 10 to the power of negative 6 to avoid numerical explosion caused by the denominator approaching zero during the calculation process, while not affecting the normal calculation results.

[0122] Preferably, the comprehensive multidimensional time-series data is input into a deep spatiotemporal information extraction network, and a particle swarm optimization algorithm is used to globally optimize the hyperparameters of the deep spatiotemporal information extraction network, including:

[0123] The deep spatiotemporal information extraction network includes a dilated one-dimensional convolutional layer, a bidirectional gated recurrent layer, a single-head attention layer, and a linear classification layer.

[0124] The integrated multidimensional time series data Inputting the dilated one-dimensional convolutional layer and the bidirectional gated recurrent layer, the sequence feature representation is extracted. ;

[0125] The sequence feature representation is input into the single-head attention layer for weighted aggregation to obtain the spatiotemporal embedding feature vector. Its expression is:

[0126]

[0127]

[0128] in, Indicates attention weights. Represents the attention vector. The temporal window length is represented; the spatiotemporal embedding feature vector is input into the linear classification layer and processed by a normalized exponential function to output the rating prediction probability. The classification cross-entropy loss is calculated based on the rating prediction probability. Its expression is:

[0129]

[0130]

[0131]

[0132] in, This represents the output of the linear classification layer. and These represent the classification weight matrix and the classification bias vector, respectively. Indicates the true rating label;

[0133] Calculate the spatiotemporal embedding feature vector relative to the device conditional parameters. The partial derivative Jacobian matrix Calculate the partial derivative Jacobian matrix and the equally probable positive and negative one random vectors. The square of the L2 norm of the product yields the structure sensitivity regularization term. Its expression is:

[0134]

[0135]

[0136] Constructing a system containing learning rates The timing window length Network Hidden Layer Dimension Random inactivation rate and structural sensitivity weights particle position Its expression is:

[0137]

[0138] At the set number of training steps, the individual fitness is calculated based on the classification cross-entropy loss and the structure sensitivity regularization term. The particle swarm optimization algorithm updates the particle positions by using the individual fitness to achieve global optimization, and its expression is:

[0139]

[0140] Sequence feature representation is a set of intermediate feature vectors containing spatiotemporal dynamic information extracted from multidimensional time-series data after processing by a dilated one-dimensional convolutional layer and a bidirectional gated recurrent layer.

[0141] Attention vectors are trainable vectors used to calculate the weights of sequence feature representations. They are preferably 128-dimensional vectors with the same dimension as the sequence feature representations to adapt to the feature dimension and ensure the effectiveness of attention weighted aggregation, while taking into account both computational efficiency and representational ability.

[0142] Attention weights are normalized weights calculated from attention vectors and sequence feature representations, and are used to highlight key temporal segments for evaluating flight operation quality.

[0143] The time window length is the length of the time segment from which comprehensive multi-dimensional time series data is extracted for network input. It is preferably an integer between 200 and 1200 to cover the effective memory scale of manipulation-response and adapt to the differences in time characteristics caused by different simulator transmission delays and motion cueing algorithms.

[0144] Spatiotemporal embedding feature vectors are low-dimensional, compact feature vectors obtained by attention-weighted aggregation, integrating key spatiotemporal information from comprehensive multi-dimensional time series data.

[0145] The weight matrix of the linear classification layer is a trainable matrix used for feature mapping in the linear classification layer. It is preferably a randomly initialized matrix of 128 rows and 3 columns. The spatiotemporal embedding feature vector is 128-dimensional, and the output is a 3-class rating result to ensure dimension matching.

[0146] The bias vector of the linear classification layer is a trainable vector used for bias adjustment in the linear classification layer. It is preferably a 3-dimensional zero-initialization vector, corresponding to the 3-class rating output. During initialization, it is necessary to avoid introducing additional bias interference to the training.

[0147] The output of the classification layer is the original output value obtained by the linear classification layer after performing a linear transformation on the spatiotemporal embedded feature vector.

[0148] The rating prediction probability is the probability distribution obtained by processing the output of the classification layer through a normalized exponential function. The value ranges from 0 to 1, and the sum is 1.

[0149] The real rating label is a rating result [collection parameter] used to supervise the actual flight operation quality rating of network training. It can be collected through manual rating results or standardized assessment conclusions entered into the training management system.

[0150] Classification cross-entropy loss is a loss value calculated based on the predicted rating probability and the actual rating label, used to measure the difference between the network's prediction and the actual situation.

[0151] The Jacobian matrix is ​​a matrix composed of the partial derivatives of the spatiotemporal embedded feature vectors with respect to device condition parameters, and is used to quantify the sensitivity of features to device structural conditions.

[0152] An equally probable positive and negative one random vector is a vector in which the elements are randomly selected as positive or negative one with equal probability. It is preferably a 6-dimensional vector because the device condition parameter is 6-dimensional, which ensures that the dimensions of the matrix and vector product calculation match.

[0153] The structural sensitivity regularization term is a regularization term obtained by approximating the square of the second norm of the product of the Jacobian matrix and the equally probable positive and negative one random vectors. It is used to suppress the excessive dependence of the spatiotemporal embedded feature vector on the device condition parameters.

[0154] Particle positions are vectors encoded from the hyperparameters of the deep network, used for global optimization in the particle swarm optimization algorithm.

[0155] The learning rate is the step size for updating parameters during deep network training. It is preferably between 10 to the power of -5 and 5 multiplied by 10 to the power of -3 to adapt to the non-stationary nature of time-series data and balance training stability and convergence speed.

[0156] The hidden layer dimension of the network is the output feature dimension of the bidirectional gated recurrent layer and the dilated one-dimensional convolutional layer. It is preferably an integer between 64 and 512 to balance the model's representational power and computational cost, and to adapt to the complexity of time series data from the flight simulator.

[0157] The random inactivation rate is the proportion of neurons randomly discarded during deep network training, preferably between 0 and 0.5, to suppress model overfitting and retain the ability to extract key features.

[0158] The structural sensitivity weight is the weight coefficient of the structural sensitivity regularization term in the total loss, preferably between 0 and 1, in order to balance the classification loss and the structural sensitivity regularization term, and avoid excessive suppression of structural information or failure to effectively isolate structural interference.

[0159] The set number of training steps is the number of times the network training gradient is updated when each particle evaluates its fitness in the particle swarm optimization algorithm. It is preferably 300 steps to ensure optimization efficiency by having short training and evaluation, while avoiding inaccurate fitness evaluation caused by insufficient training.

[0160] The individual particle fitness is the objective function value calculated based on the classification cross-entropy loss and the structure sensitivity regularization term, which is used to drive the particle swarm algorithm to update the particle position.

[0161] In detail, the deep spatiotemporal information extraction network adopts a combined structure of dilated one-dimensional convolutional layers and bidirectional gated recurrent layers. This is because dilated convolution can expand the receptive field without increasing computational cost, capturing long-term temporal dependencies, while bidirectional gated recurrent layers can effectively model the forward and backward dynamic correlations of control and response. The combination of these two layers adapts to the characteristics of flight simulator time-series data, such as the ability to fully extract temporal evolution features of continuous control adjustments and attitude responses in instrument approach maneuvers. The design of a single-head attention layer can automatically focus on key operational segments, such as significantly increasing the feature weights corresponding to stick correction actions during landing, thus enhancing the relevance of the evaluation. The structural sensitivity regularization term is approximated by the squared L2 of the product of the Jacobian matrix and equally probable positive and negative one-order random vectors, avoiding the huge computational burden of explicitly constructing a high-dimensional Jacobian matrix. The regularization term effectively quantifies the sensitivity of features to device conditions. For example, when device transmission delay changes, the regularization term constrains the change in the spatiotemporal embedded feature vector. The five hyperparameters of particle position encoding are core parameters for deep network training and are strongly coupled with simulator structural features. The learning rate adapts to data non-stationarity, the temporal window length matches the delay scale, the network hidden layer dimension adapts to feature complexity, the random inactivation rate suppresses overfitting, and the structural sensitivity weight balances classification and anti-interference objectives. The individual fitness function weights the classification cross-entropy loss and the structural sensitivity regularization term, enabling the network to actively suppress dependence on device structural conditions while learning accurate ratings. For example, when training with multiple devices, the model will not be biased towards learning the structural features of a particular device due to its high data proportion.

[0162] In detail, three dilated one-dimensional convolutional layers are set, each with a kernel size of 5 and dilation rates of 1, 2, and 4 respectively. The number of output channels is consistent with the dimension of the hidden layers. The activation function is a linear rectified function, and a batch normalization layer is added after each layer. A single bidirectional gated recurrent layer is set, with half the number of hidden units of the hidden layers. The outputs of the forward and backward hidden units are concatenated to obtain features consistent with the dimension of the hidden layers. Gradient clipping is used to prevent gradient explosion, with a clipping threshold of 1.0. The attention vector uses the Hessian initialization method to ensure that the initial values ​​can evenly distribute the attention weights. It is updated along with the network parameters during training. Equal probability positive and negative one random vectors are generated by a random number generator. Each element independently takes the value 1 or -1 with a 50% probability, and is regenerated once for each training batch. The basic principles of particle swarm optimization are... The parameters were set to 20 particles, 30 iterations, a learning factor of 2.0, and an initial inertia weight of 0.7. The deep network training employed an adaptive momentum estimation optimizer with a momentum parameter of 0.9, a second-order momentum parameter of 0.999, and a weight decay coefficient of 10 to the power of negative 4. Random inactivation was applied only to the outputs of the dilated one-dimensional convolutional layer and the bidirectional gated recurrent layer; random inactivation was not enabled for the attention layer and the linear classification layer. The true rating labels used one-hot encoding, with the three ratings corresponding to [1,0,0], [0,1,0], and [0,0,1], respectively. The data was automatically converted to this encoding format during training management system input. The training steps were set to 300 steps, with a fixed batch size of 64 for each step. Data was completely divided according to a single training session to avoid training interference caused by cross-session data mixing.

[0163] Preferably, in the global optimization process, extracting the set of individual optimal positions of the particle swarm optimization algorithm to calculate the particle swarm discrete coefficients includes:

[0164] Obtain the optimal position of each particle in the set of optimal individual positions. ;

[0165] Based on the set lower bound of the search space Upper bound of search space The optimal positions of each individual particle are then normalized to obtain the optimal positions of each normalized individual particle. Its expression is:

[0166]

[0167] Calculate the covariance matrix formed by the optimal positions of each normalized individual. Its expression is:

[0168]

[0169] in, This represents the total number of particles in the particle swarm optimization algorithm. This represents the function for calculating covariance. Indicates the first The normalized individual optimal position vector of each particle;

[0170] Calculate the largest eigenvalue of the covariance matrix. With the trace of the covariance matrix The ratio of these values ​​is used as the particle swarm discrete coefficient. Its expression is:

[0171]

[0172] The optimal position of an individual particle is the hyperparameter combination vector that each particle finds during the iteration process in the particle swarm optimization algorithm, which maximizes the individual's fitness.

[0173] The lower bound of the search space is the minimum boundary value allowed for each hyperparameter in the particle swarm optimization algorithm. The preferred values ​​are a learning rate of 10 to the power of negative 5, a temporal window length of 200, a network hidden layer dimension of 64, a random inactivation rate of 0, and a structure sensitivity weight of 0, to ensure that the optimization does not exceed the effective range.

[0174] The upper bound of the search space is the maximum boundary value allowed for each hyperparameter in the particle swarm optimization algorithm. The preferred values ​​are a learning rate of 5 times 10 to the power of negative cube, a temporal window length of 1200, a network hidden layer dimension of 512, a random inactivation rate of 0.5, and a structure sensitivity weight of 1, which are consistent with the lower bound of the search space, balancing model performance and computational efficiency.

[0175] The normalized individual optimal position is a standardized vector obtained by linearly normalizing the optimal position of an individual particle according to the upper and lower bounds of the search space. It is used to eliminate the differences in the dimensions of different hyperparameters.

[0176] The total number of particles is the number of particles participating in the global optimization of the particle swarm optimization algorithm. It is preferably 20 to balance the diversity of optimization and computational efficiency, and to avoid the optimization getting stuck in local optima or excessively increasing the computational burden due to too few particles.

[0177] The covariance matrix is ​​a square matrix composed of the normalized individual optimal positions of all particles, used to characterize the dispersion and correlation of the particle swarm optimization distribution.

[0178] The largest eigenvalue of the covariance matrix is ​​the eigenvalue with the largest numerical value among the eigenvalues ​​of the covariance matrix, and it is used to reflect the intensity of the main discrete directions of the particle swarm optimization distribution.

[0179] The trace of the covariance matrix is ​​the sum of the elements on the main diagonal of the covariance matrix, and is used to characterize the overall discreteness of the particle swarm optimization distribution.

[0180] The particle swarm discrete coefficient is the ratio of the largest eigenvalue of the covariance matrix to the trace of the covariance matrix. It is used to quantify the degree of splitting in the hyperparameter optimization distribution and its value ranges from 0 to 1.

[0181] In detail, boundary normalization is performed on the optimal positions of individual particles because the ranges and dimensions of different hyperparameters vary greatly. For example, the learning rate ranges from 10⁻⁵ to 5 × 10⁻³, while the time window length ranges from 200 to 1200. Directly calculating the covariance matrix would lead to distorted results due to the imbalance of dimensions. After normalization, all hyperparameters are mapped to the interval between 0 and 1, ensuring a balanced contribution of each dimension to the covariance matrix. The particle swarm dispersion coefficient is the ratio of the largest eigenvalue of the covariance matrix to the trace, capturing the geometric shape of the particle swarm optimization distribution. When hyperparameter optimization is induced by device condition parameters to split, the particle swarm will converge around multiple local optima. This leads to a significant increase in the largest eigenvalue of the covariance matrix, with the coefficient of variation approaching 1, while the coefficient of variation approaches 0 when the optimization distribution is concentrated. The covariance matrix is ​​constructed by choosing the set of individual optimal positions rather than the global optimal positions because the individual optimal positions can completely preserve the optimization trajectory and distribution characteristics of the particle swarm. The global optimal positions only reflect a single optimal solution and cannot reflect the splitting of the overall optimization distribution. For example, in multi-device mixed training, some particles converge around the hyperparameters of the adaptation device A, while some converge around the hyperparameters of the adaptation device B. The covariance matrix of the set of individual optimal positions will clearly show this splitting, while the global optimal position is only the optimal solution corresponding to one device and cannot reflect the distribution difference.

[0182] In detail, the specific values ​​for the lower and upper bounds of the search space for each dimension of particle position are as follows: learning rate: lower bound 10^-5, upper bound 5 * 10^-3; temporal window length: lower bound 200, upper bound 1200; network hidden layer dimension: lower bound 64, upper bound 512; random inactivation rate: lower bound 0, upper bound 0.5; structure sensitivity weight: lower bound 0, upper bound 1. All boundary values ​​are determined based on the characteristics of time-series data from the flight simulator and deep network training experience. The total number of particles is fixed at 20. This value has balanced optimization efficiency and diversity in multiple engineering verifications, avoiding insufficient optimization coverage due to fewer than 10 particles and avoiding problems with more than 30 particles. The computational cost increases dramatically due to the particle's position. The covariance matrix has the same dimensionality as the particle position, being 5-dimensional, corresponding to five hyperparameters: learning rate, temporal window length, network hidden layer dimension, random inactivation rate, and structure sensitivity weights. The eigenvalues ​​are solved using the singular value decomposition algorithm, which is highly stable in numerical computation and can accurately extract the eigenvalues ​​of the covariance matrix. The optimal position of an individual particle is updated once per iteration. When the current particle's fitness is better than the historical best fitness, the current position replaces the historical optimal position. The initial optimal position of the individual particle is the initial position of the particle, ensuring that the optimal position of the individual particle can track the particle's optimization progress in real time.

[0183] Preferably, the inertia weight of the particle swarm optimization algorithm and the divergence cost weight in the preset optimization objective function are continuously modulated based on the particle swarm discrete coefficients to suppress the dispersion of the hyperparameter optimization distribution induced by the device condition parameters, including:

[0184] Based on the zero constant The particle swarm discrete coefficients Boundary constraint processing is performed to obtain the discrete coefficients of the constrained particle swarm. Its expression is:

[0185]

[0186] Calculate the difference between the constraint particle swarm discrete coefficients and one, and add the zero-prevention constant to obtain the denominator term; calculate the ratio of the constraint particle swarm discrete coefficients to the denominator term and add one, then take the natural logarithm of the summation result to obtain the divergence cost function value. Its expression is:

[0187]

[0188] Calculate the discrete coefficients and center mapping parameters of the constrained particle swarm. The difference is compared with the shape adjustment parameter. After multiplication, perform a sigmoid mapping function. Process the mapping and then correlate the mapping results with the maximum divergence cost weights. Multiply to obtain the divergence cost weight. Its expression is:

[0189]

[0190] Calculate the difference between the discrete coefficients of the constrained particle swarm and the maximum inertia weight. and minimum inertia weight The differences are multiplied, and the product is added to the minimum inertia weight to obtain the inertia weight. Its expression is:

[0191]

[0192] The classification cross-entropy loss The structural sensitivity weights With the structure sensitivity regularization term The product of, and the divergence cost weight With the value of the divergence cost function The products of these factors are summed to update the individual fitness. Its expression is:

[0193]

[0194] The constrained particle swarm discrete coefficients are standardized coefficients obtained by applying zero-constant boundary constraints to the particle swarm discrete coefficients, and are used to avoid computational singularities.

[0195] The divergence cost function value is a cost quantification value calculated by constraining the particle swarm discrete coefficients, and is used to measure the severity of the splitting of the particle swarm optimization distribution.

[0196] The center mapping parameter is the central reference value of the S-shaped mapping function, preferably 0.7, which is the key threshold range of the particle swarm discrete coefficient, enabling the mapping function to quickly adjust the weights when the risk of splitting is significant (the coefficient is close to 0.7).

[0197] The shape adjustment parameter is a coefficient that adjusts the steepness of the S-shaped mapping function curve. It is preferably 10 to balance the sensitivity and smoothness of the mapping and avoid oscillations in the optimization caused by sudden weight changes.

[0198] The S-shaped mapping function is a nonlinear function that maps the discrete coefficients of a constrained particle swarm to the interval between 0 and 1. It is preferably a standard logic function and is a commonly used smooth nonlinear mapping tool in engineering, which can realize continuous and gradual weight adjustment.

[0199] The maximum divergence cost weight is the maximum weight upper limit of the divergence cost function value, preferably 0.1, to avoid the excessive divergence cost weight suppressing the classification loss and to ensure that the model takes into account both rating accuracy and anti-split ability.

[0200] The divergence cost weight is a cost weight obtained by dynamically modulating the discrete coefficients of the constrained particle swarm, and is used to quantify the impact of splitting risk on particle fitness.

[0201] The maximum inertia weight is the maximum value of the inertia weight in the particle swarm algorithm, preferably 0.9, to ensure that the particle swarm has sufficient global exploration capability in the early stage.

[0202] The minimum inertia weight is the minimum value of the inertia weight in the particle swarm algorithm, preferably 0.4, to avoid excessive inertia in the later stages of the particle swarm algorithm causing local optima to stagnate.

[0203] Inertia weight is a particle swarm optimization parameter obtained by dynamically modulating the discrete coefficients of the constrained particle swarm, used to balance the exploration and exploitation capabilities of particles.

[0204] The iteration number is the current iteration number of the particle swarm optimization algorithm, used to mark the optimization process.

[0205] The updated individual fitness is the objective function value that integrates the weighted sum of classification cross-entropy loss, structure sensitivity regularization term and divergence cost, and is used to drive the particle swarm algorithm to update particle positions.

[0206] Using a zero-constant to constrain the particle swarm discrete coefficients is to avoid computational singularities where the denominator is zero when the discrete coefficients approach 0 or 1, ensuring the numerical stability of the modulation process. For example, when the discrete coefficient is 0, it becomes 10 to the power of negative 6 after constraint. The logarithmic form of the divergence cost function design can nonlinearly amplify the discrete coefficients in the 0 to 1 interval. When the discrete coefficient increases from 0.5 to 0.9, the cost function value increases from 0.81 to 2.94, significantly amplifying the penalty for severe splits, while the penalty is slight when the coefficient is below 0.5, achieving differentiated risk management. The S-shaped mapping function modulates the divergence cost weights, making the weights change smoothly and gradually with the discrete coefficients, avoiding optimization oscillations caused by threshold-like abrupt changes. For example, the weight is 0.025 when the discrete coefficient is 0.6, and 0 when it is 0.8. 075 achieves a precise match between splitting risk and penalty intensity; the rule of linearly modulating inertia weights ensures that the larger the dispersion coefficient (the more severe the split), the smaller the inertia weight, gradually decreasing from 0.9 to 0.4, reducing global exploration by particles and prompting them to converge towards a concentrated region. For example, the inertia weight is 0.44 when the dispersion coefficient is 0.9 and 0.78 when the dispersion coefficient is 0.3, balancing exploration and development; the fitness update formula with a three-part weighted sum integrates classification loss, structural anti-interference loss, and splitting risk loss, enabling the particle swarm to simultaneously consider rating accuracy, device independence, and distribution concentration during optimization. For example, when multi-device mixed training leads to a dispersion coefficient of 0.85, the divergence cost weight increases, and the particle fitness is significantly affected by the splitting penalty, driving particles to converge towards a unified optimal hyperparameter region.

[0207] In detail, the S-shaped mapping function adopts a standard logic function, and the expression is that the output equals 1 divided by 1 plus the negative power of the natural exponent. The input is the shape adjustment parameter multiplied by the difference between the constraint particle swarm discrete coefficient and the center mapping parameter. The output range is stable between 0 and 1. The values ​​of the center mapping parameter 0.7, shape adjustment parameter 10, maximum divergence cost weight 0.1, maximum inertia weight 0.9, and minimum inertia weight 0.4 can adapt to the optimization needs of different device combinations and training scenarios. The inertia weight is updated once per iteration, synchronized with the particle position update of the particle swarm algorithm, ensuring that the modulation responds promptly to changes in the optimization distribution. The maximum number of iterations is 30, consistent with the total number of iterations of the particle swarm algorithm, covering the complete optimization process. The divergence cost weight has no additional iteration decay or enhancement rules and is completely determined by the current generation of constraint particle swarm discrete coefficients, ensuring the real-time performance and objectivity of the modulation. In the calculation of the updated individual fitness, the weight ratio of each part of the loss is naturally balanced through dynamic modulation, without the need to set additional fixed ratio coefficients, avoiding the implementation complexity caused by multiple parameter adjustments.

[0208] Preferably, the deep spatiotemporal information extraction network is trained based on the optimal hyperparameters obtained after modulation convergence, so as to perform sliding window inference on the assessment sequence and output rating labels, including:

[0209] Extract the global optimum position after the particle swarm optimization algorithm converges. The globally optimal position is decoded to obtain the result containing the optimal timing window length. The optimal hyperparameters, including those mentioned above;

[0210] Based on the student identity identifier With the subject identification A training set is constructed by performing hash partitioning, and the training of the deep spatiotemporal information extraction network is completed using the optimal hyperparameters.

[0211] For a total number of sampling points The assessment sequence The optimal timing window length is used as the sliding window length, and half of the optimal timing window length is rounded down as the sliding window step size. The assessment sequence is divided into A sequence window, whose expression is:

[0212]

[0213]

[0214] in, This indicates a round-down operation;

[0215] Each sequence window is input into the trained deep spatiotemporal information extraction network, which outputs a corresponding window classification feature vector. ;

[0216] The global average feature vector is obtained by summing and averaging the window classification feature vectors of all the sequence windows. Its expression is:

[0217]

[0218] The global average feature vector is processed by a normalized exponential function to obtain the global classification probability distribution. Its expression is:

[0219]

[0220] in, This represents the normalized exponential function;

[0221] Extract the category index corresponding to the maximum probability value in the global classification probability distribution. Its expression is:

[0222]

[0223] in, In the global classification probability distribution, the first... The probability value of the class;

[0224] The category index is converted into a rating label based on a preset rating mapping dictionary, indicating that the user is qualified for real device access, qualified but requires additional practice, or unqualified and forced to retrain.

[0225] The global optimal position is the hyperparameter combination vector that achieves the best global fitness among all the optimal positions of individual particles after the particle swarm optimization algorithm has converged iteratively.

[0226] The optimal hyperparameters are the set of deep network training parameters obtained after decoding the global optimal position, including the optimal learning rate, optimal temporal window length, optimal network hidden layer dimension, optimal random inactivation rate, and optimal structure sensitivity weights.

[0227] The optimal time window length is the window length used to extract time series data in the optimal hyperparameters. It is obtained by decoding the global optimal position and is an effective memory scale adapted to the manipulation-response relationship.

[0228] The total number of sampling points in the assessment sequence is the total number of data points after interpolation at a uniform sampling frequency, obtained by multiplying the assessment duration by the uniform sampling frequency.

[0229] The sliding window step size is the time interval between two adjacent sequence windows during sliding window inference, which is obtained by rounding down half of the optimal time window length.

[0230] The sequence window number is the total number of windows obtained after the test sequence is divided by sliding window, which is calculated from the total number of sampling points of the test sequence, the optimal time window length, and the sliding window step size.

[0231] A sequence window is a local time series data segment obtained by truncating the test sequence according to the optimal time series window length and sliding window step size.

[0232] The window classification feature vector is the original feature vector output by the linear classification layer after a deep network has been trained with a single sequence window as input.

[0233] The global average feature vector is a comprehensive feature vector obtained by summing and averaging the classification feature vectors of all sequence windows. It is used to smooth out the anomalous effects of local time segments.

[0234] The global classification probability distribution is the three-class rating probability distribution obtained by processing the global average feature vector with a normalized exponential function. The values ​​range from 0 to 1, and the sum is 1.

[0235] The category index is an integer identifier corresponding to the maximum probability value in the global classification probability distribution, with a value of 1, 2, or 3, corresponding to the three categories of rating results respectively.

[0236] The rating mapping dictionary is a set of mapping rules that convert category indexes into business-oriented rating labels. The preferred values ​​are 1 for those who are qualified to enter the real aircraft, 2 for those who are qualified but need to practice more, and 3 for those who are unqualified and must retrain. The values ​​are selected based on the actual business needs of flight school training graduation assessments and cover rating scenarios for students with different ability levels.

[0237] In detail, the training set is constructed based on hash partitioning of student and subject identifiers. The core principle is to achieve leak-free data partitioning through hash operations on unique identifiers, preventing data from the same student and subject from being mixed across the training and validation sets, thus ensuring the objectivity of the model's generalization evaluation. For example, if the student identifier is 1001 and the subject identifier is 05, after hashing and taking the modulo 20, the result is 12, which is then included in the training set. The sliding window inference rule uses the optimal temporal window length as the sliding window length, with half of that length rounded down as the step size. This ensures that each local temporal segment is covered while maintaining the continuity of temporal features through a 50% overlap rate. For example, if the optimal temporal window length is 400 sampling points, the sliding window step size is 200 sampling points, and adjacent windows overlap by 200 sampling points. The method of summing and averaging the feature vectors of window classification can effectively smooth the impact of local abnormal manipulation segments on the rating results. For example, if the feature vectors of some windows in a certain assessment sequence deviate due to instantaneous operational errors, they will not excessively affect the final rating after global averaging. The three-level rating mapping dictionary directly connects to the graduation assessment requirements of flight schools, converting the category index output by the algorithm into rating conclusions that can be directly used by business personnel. The optimal hyperparameters are obtained by decoding the global optimal position and then used for full training, realizing a closed loop between particle swarm optimization and deep network training, ensuring that the model can make full use of the optimal configuration obtained by optimization. The sliding window step size is fixed at half the length of the optimal temporal window and rounded down to ensure the consistency of the overlap rate under different window lengths and avoid fluctuations in inference results caused by arbitrary step size settings.

[0238] In detail, the hash partitioning uses the Secure Hash Algorithm 256 (SHA-256). The first 8 bytes of the hash result are converted to an integer and then modulo 20. Modulo values ​​0 to 13 are assigned to the training set (70%), 14 to 16 to the validation set (15%), and 17 to 19 to the test set (15%), ensuring the randomness and lack of data leakage in the partitioning. The parameters for full training of the deep network are set as follows: 30 training epochs, batch size 64, learning rate using the optimal hyperparameter decoding value, weight decay coefficient of 10 to the power of -4, and early stopping rule: training stops if the validation set loss does not decrease for 5 consecutive epochs, ensuring sufficient training and preventing overfitting. During sliding window inference, if the assessment sequence length is less than the optimal time window length, zero-padding is used to pad the sequence length to the optimal time window length, with the zero-padding position at the end of the sequence to avoid feature loss due to truncation. The rating mapping dictionary supports expansion, with the expansion rule that new category indexes must be added according to ability level. The allocation is done in ascending order, and newly added rating labels must correspond to the flight school assessment standards to ensure consistency of the expanded mapping rules. The decoding rule for the global optimal position is that hyperparameters encoded in the logarithmic domain (learning rate, temporal window length, network hidden layer dimension) are converted to their original values ​​through exponential operations, while hyperparameters encoded in the linear domain (random inactivation rate, structure sensitivity weights) are directly taken from their original values. For example, the logarithmic domain learning rate is decoded into the result of exponential operations. The sliding window overlap rule is that the overlapping part of adjacent windows is taken as the latter part of the previous window and the former part of the next window to ensure the continuity of temporal data. The full training uses an adaptive momentum estimation optimizer with a momentum parameter of 0.9 and a second-order momentum parameter of 0.999, consistent with the optimizer in the particle swarm optimization stage. The evaluation indicators for model performance include accuracy, precision, recall, and F1 score. The passing standard is an accuracy of not less than 90% and an F1 score of not less than 0.85 to ensure that the model meets the accuracy requirements of practical applications.

[0239] like Figure 2 As shown, Figure 2 This diagram illustrates the parameter variation trends during the iterative process of the particle swarm optimization (PSO) algorithm. The horizontal axis represents the number of iterations (1 to 30), the left vertical axis represents the coefficient of variation (SBC), which characterizes the degree of splitting in the hyperparameter optimization distribution, and the right vertical axis represents the weights (inertia weight and divergence cost weight). The SBC monotonically decreases from approximately 0.92 to approximately 0.55 with increasing iterations, indicating a gradual reduction in the degree of splitting in the hyperparameter optimization distribution. The inertia weight monotonically increases from approximately 0.05 to approximately 0.55, reflecting the adaptive adjustment of the PSO algorithm from an exploration-oriented focus to an exploration-oriented focus as the SBC decreases. The divergence cost weight fluctuates slightly in the early stages of iteration before stabilizing at a low level of approximately 0.05, reflecting that the penalty weight for optimization splitting remains within a small and stable range due to the decrease in the SBC. These three factors change synergistically, intuitively demonstrating the adaptive modulation mechanism based on the SBC, which drives the gradual convergence of the hyperparameter optimization distribution and suppresses hyperparameter splitting induced by device conditions.

[0240] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A deep learning-based method for evaluating the flight operation quality of an aviation simulator, characterized in that, include: Acquire the flight simulator's control input data, attitude response data, simulator condition metadata, and assessment sequences; Calculate the transmission delay between the manipulation input data and the attitude response data, and construct device condition parameters based on the transmission delay and the simulator condition metadata; Delay mapping correction is performed on the attitude response data based on the transmission delay to obtain corrected attitude data; control impulse degree and response sensitivity are calculated based on the manipulation input data and the corrected attitude data; the manipulation input data, the corrected attitude data, the control impulse degree, the response sensitivity, and the device condition parameters are concatenated to construct comprehensive multidimensional time series data; The comprehensive multidimensional time series data is input into the deep spatiotemporal information extraction network, and the hyperparameters of the deep spatiotemporal information extraction network are globally optimized using the particle swarm optimization algorithm. During the global optimization process, the set of individual optimal positions of the particle swarm optimization algorithm is extracted to calculate the particle swarm discrete coefficients; The inertia weight of the particle swarm algorithm and the divergence cost weight in the preset optimization objective function are continuously modulated based on the particle swarm discrete coefficients to suppress the optimization distribution discrepancy of the hyperparameters induced by the device condition parameters. The deep spatiotemporal information extraction network is trained based on the optimal hyperparameters obtained after modulation convergence, so as to perform sliding window inference on the assessment sequence and output rating labels.

2. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 1, characterized in that, Acquire the flight simulator's control input data, attitude response data, simulator condition metadata, and assessment sequences, including: The lateral displacement of the joystick, the longitudinal displacement of the joystick, the deflection of the pedals, and the throttle status, all with control input timestamps, are acquired as the control input data. The attitude angle, angular velocity, airspeed, and altitude, which include the instrument display refresh timestamp, visual frame timestamp, and motion command output timestamp, are obtained as the attitude response data. The instructor support level, situation complexity level, device identifier, student identity identifier, and subject identity identifier are obtained as the simulator condition metadata. Obtain the timing data of the manipulation and attitude to be evaluated, and use it as the assessment sequence.

3. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 2, characterized in that, Calculate the transmission delay between the manipulation input data and the attitude response data, and construct device condition parameters based on the transmission delay and the simulator condition metadata, including: The differences between the instrument display refresh timestamp, the visual frame timestamp, the motion command output timestamp, and the control input timestamp are calculated respectively to obtain the instrument channel delay sequence, the visual channel delay sequence, and the motion channel delay sequence. The medians of the instrument channel delay sequence, the visual channel delay sequence, and the motion channel delay sequence are extracted respectively and used as the instrument robust transmission delay, the visual robust transmission delay, and the motion robust transmission delay, respectively; if there is no motion command output timestamp, the visual robust transmission delay is used as the motion robust transmission delay; The device identifiers are normalized based on the total number of devices to obtain device code identifiers; The device condition parameters are constructed by combining the instrument robust transmission delay, the vision robust transmission delay, the motion robust transmission delay, the instructor support level, the context complexity level, and the device coding identifier.

4. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 3, characterized in that, Delay mapping correction is performed on the attitude response data based on the transmission delay to obtain corrected attitude data; control impulse degree and response sensitivity are calculated based on the manipulation input data and the corrected attitude data; the manipulation input data, the corrected attitude data, the control impulse degree, the response sensitivity, and the device condition parameters are concatenated to construct comprehensive multi-dimensional time-series data, including: The manipulation input data and the attitude response data are interpolated using a uniform sampling frequency to obtain equally spaced manipulation input data and equally spaced attitude response data. Based on the ratio of the instrument's robust transmission delay to the sampling step size corresponding to the unified sampling frequency, the time axis of the equally spaced attitude response data is shifted to obtain the corrected attitude data; The equally spaced manipulation input data and the corrected attitude data are filtered and normalized respectively to obtain normalized manipulation input data and normalized corrected attitude data. The normalized manipulation input data is differentially processed to obtain the derivative of the manipulation input data; The control impulse degree is obtained by calculating the ratio of the sum of the first norm of the derivatives of the manipulation input data to the sum of the second norm of the normalized manipulation input data. The ratio of the differential change in the normalized corrected attitude data to the differential change in the normalized manipulation input data is calculated and smoothed to obtain the response sensitivity. The device condition parameters are linearly mapped and tiled to obtain the tiled condition parameters; The normalized manipulation input data, the derivative of the manipulation input data, the normalized corrected attitude data, the response sensitivity, the control impulse degree, and the tiling condition parameters are spliced ​​together to construct the comprehensive multidimensional time series data.

5. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 4, characterized in that, The comprehensive multidimensional time-series data is input into a deep spatiotemporal information extraction network, and the hyperparameters of the deep spatiotemporal information extraction network are globally optimized using a particle swarm optimization algorithm, including: The deep spatiotemporal information extraction network includes a dilated one-dimensional convolutional layer, a bidirectional gated recurrent layer, a single-head attention layer, and a linear classification layer. The comprehensive multidimensional time series data is input into the dilated one-dimensional convolutional layer and the bidirectional gated recurrent layer to extract sequence feature representations. The sequence feature representation is input into the single-head attention layer for weighted aggregation to obtain the spatiotemporal embedding feature vector; The spatiotemporal embedding feature vector is input into the linear classification layer and processed by the normalized exponential function to output the rating prediction probability. The classification cross-entropy loss is calculated based on the rating prediction probability. Calculate the partial derivative Jacobian matrix of the spatiotemporal embedding feature vector with respect to the device condition parameter, and calculate the square of the second norm of the product of the partial derivative Jacobian matrix and the equally probable positive and negative one random vectors to obtain the structural sensitivity regularization term; Construct particle positions that include learning rate, temporal window length, network hidden layer dimension, random inactivation rate, and structure sensitivity weights; At a set number of training steps, the individual fitness is calculated based on the classification cross-entropy loss and the structure sensitivity regularization term, and the particle swarm algorithm is used to update the particle position to achieve global optimization.

6. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 5, characterized in that, In the global optimization process, the set of individual optimal positions of the particle swarm optimization algorithm is extracted to calculate the particle swarm discrete coefficients, including: Obtain the optimal position of each individual particle in the set of optimal individual positions; Based on the set lower and upper bounds of the search space, the optimal positions of each individual particle are normalized to obtain the optimal positions of each normalized individual particle. Calculate the covariance matrix formed by the optimal positions of each normalized individual; The ratio of the largest eigenvalue of the covariance matrix to the trace of the covariance matrix is ​​calculated and used as the particle swarm discrete coefficient.

7. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 6, characterized in that, Based on the particle swarm discrete coefficients, the inertia weights of the particle swarm algorithm and the divergence cost weights in the preset optimization objective function are continuously modulated to suppress the dispersion of the hyperparameter optimization distribution induced by the device condition parameters, including: The particle swarm discrete coefficients are subjected to boundary constraint processing based on the zero-prevention constant to obtain the constrained particle swarm discrete coefficients. Calculate the difference between the constraint particle swarm discrete coefficient and one, and add the zero-prevention constant to obtain the denominator term; calculate the ratio of the constraint particle swarm discrete coefficient to the denominator term and add one, then take the natural logarithm of the summation result to obtain the divergence cost function value. The difference between the discrete coefficients of the constrained particle swarm and the center mapping parameter is calculated. The difference is multiplied by the shape adjustment parameter and then processed by the S-shaped mapping function. The mapping result is multiplied by the maximum divergence cost weight to obtain the divergence cost weight. Calculate the difference between the discrete coefficients of the constrained particle swarm, multiply the difference between the maximum inertia weight and the minimum inertia weight, and add the product to the minimum inertia weight to obtain the inertia weight; The individual fitness is updated by summing the product of the classification cross-entropy loss, the product of the structure sensitivity weight and the structure sensitivity regularization term, and the product of the divergence cost weight and the divergence cost function value.

8. The method for evaluating the flight operation quality of an aviation simulator based on deep learning according to claim 7, characterized in that, The deep spatiotemporal information extraction network is trained based on the optimal hyperparameters obtained after modulation convergence, and then used for sliding window inference on the assessment sequence to output rating labels, including: Extract the global optimal position after the particle swarm optimization algorithm converges, decode the global optimal position, and obtain the optimal hyperparameters including the optimal time window length; A training set is constructed by hash partitioning based on the student identity identifier and the subject identity identifier, and the training of the deep spatiotemporal information extraction network is completed using the optimal hyperparameters. For the assessment sequence, the optimal time window length is used as the sliding window length, and half of the optimal time window length is rounded down as the sliding window step size, so that the assessment sequence is divided into multiple sequence windows; Each sequence window is input into the trained deep spatiotemporal information extraction network, and the corresponding window classification feature vector is output. The window classification feature vectors of all the sequence windows are summed and averaged to obtain the global average feature vector; The global average feature vector is processed by a normalized exponential function to obtain the global classification probability distribution; Extract the category index corresponding to the maximum probability value in the global classification probability distribution, and convert the category index into a rating label based on a preset rating mapping dictionary.

Images