An adaptive control system for pitch vibration of an aircraft

Abstract

This invention relates to the field of aircraft technology, specifically disclosing an adaptive control system for aircraft pitch vibration. The system includes: a feature acquisition module that collects pitch vibration signals from a control stick and combines them with scene state parameters and control force parameters to complete the association and labeling of operational intentions, constructing an intention feature dataset; an intention prediction module that identifies accidental behavioral features based on the operational baseline and filters pseudo-correlated feature-intention pairs through comparison of co-occurrence patterns of multiple samples; an adaptation analysis module that distinguishes individual or common behavioral limitations through clustering and dynamic time warping algorithms, extracting and verifying misjudgment time windows; and a compensation evaluation module that determines the optimal control window for the misjudgment time window, adjusts it differently according to the type of limitation, and evaluates the optimization effect through accuracy improvement and vibration improvement. This invention is beneficial for improving the vibration control accuracy of flight simulators and can be applied to vibration control scenarios in flight simulators.

Description

Technical Field

[0001] This invention relates to the field of aircraft technology, and more specifically to an adaptive control system for aircraft pitch vibration. Background Technology

[0002] Pitch vibration control of aircraft is a core component in ensuring flight stability and operational safety. As a core device for replicating the flight state of aircraft and conducting pilot training, the realism of pitch vibration simulation and the accuracy of operational intent recognition are important prerequisites for helping pilots form standardized operating habits and improve their ability to cope with complex operating conditions.

[0003] In flight simulator training scenarios, the correlation between the pitch vibration signal of the control stick and the pilot's operational intentions, as well as differentiated vibration control for different sources of interference, have a significant impact on improving training quality and subsequent actual flight safety. Traditional control modes are not well-suited for flight simulators, making it difficult to pinpoint the time window for vibration and intention misinterpretation within the simulator, and also unable to identify the underlying limitations of misinterpretations. This results in a lack of clear direction for adjusting control strategies, hindering the continuous improvement of flight simulator training adaptability and failing to fully meet pilots' needs for realism and accuracy in simulation training.

[0004] Therefore, an adaptive control system for aircraft pitch vibration is provided. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive control system for aircraft pitch vibration to solve the aforementioned background problems.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] An adaptive control system for pitch vibration of an aircraft includes the following modules:

[0008] Feature acquisition module: used to acquire input features of joystick pitch vibration, perform operation intention association labeling on the input features, and construct an intention feature dataset after association labeling.

[0009] Intent prediction module: Extracts incidental behavioral features from the intent feature dataset, performs correlation screening between incidental behavioral features and operational intent to obtain pseudo-correlation feature-intent pairs;

[0010] The adaptation analysis module analyzes the pseudo-relevant feature-intent pairs through behavioral limitation analysis to determine whether there are behavioral limitations; if so, it performs misjudgment window identification processing on the pseudo-relevant feature-intent pairs to obtain the misjudgment time window.

[0011] Compensation evaluation module: Performs controllable evaluation processing on the misjudgment time window, determines the optimal control window, optimizes pitch vibration control based on the optimal control window for pseudo-correlation feature-intent pair, collects matching data of optimized pitch vibration and operation intent, and evaluates it.

[0012] As a further aspect of the present invention, the process of performing the associated tagging of the aforementioned operational intent is as follows:

[0013] The input features are transformed into input feature vectors, and the classification dimension of the operation intent is determined. Historical input feature vectors are obtained and combined with the classification dimension to construct the intent feature vector.

[0014] Collect and verify supplementary data, and perform matching and labeling processing on the input feature vector based on historical intent feature vectors to obtain the initially labeled operation intent;

[0015] The operation intent initially labeled is then combined with supplementary verification data for rule verification and correction to obtain the operation intent label. This label is then combined with the input feature vector to achieve associated labeling processing.

[0016] As a further aspect of the present invention, the matching tag processing is performed as follows:

[0017] A vector similarity matching algorithm is used to match and label the input feature vector with the historical intent feature vector to obtain the initially labeled operation intent.

[0018] As a further aspect of the present invention, the method for performing the correlation misjudgment screening is as follows:

[0019] Extract samples from the intent feature dataset and randomly divide them into K groups according to pilot ID;

[0020] For each incidental behavioral feature, feature-intent co-occurrence patterns are statistically analyzed in K groups. The similarity of co-occurrence patterns in K groups is calculated using the cosine similarity algorithm, and highly consistent pairs are extracted based on the similarity.

[0021] The high-consistency groups are processed for pseudo-correlation, resulting in pseudo-correlation feature-intent pairs.

[0022] As a further aspect of the present invention: the method for extracting the incidental features of the behavior is as follows:

[0023] Establish an operational baseline for each pilot based on the intent feature dataset;

[0024] The pilot's input feature vector is collected in real time, and the deviation between the input feature vector and the operating baseline is analyzed to obtain the random judgment feature.

[0025] Establish a random feature determination rule. If a random feature satisfies the random feature determination rule, then mark the current input feature vector as a random behavioral feature.

[0026] As a further aspect of the present invention: the behavioral limitation analysis is performed as follows:

[0027] Collect the frequency of occurrence of pseudo-correlation features—intent pairs—for each pilot according to their identification number, and the total frequency of pilot operations.

[0028] The individual deviation ratio is obtained by calculating the proportion of the occurrence frequency of each pilot's pseudo-correlation feature-intent pair to the total operation frequency.

[0029] The mean of the individual deviation ratios of all pilots is calculated to obtain the group deviation ratio;

[0030] The association determination is performed based on the individual deviation ratio and the group deviation ratio to obtain the result of whether the behavior is restrictive.

[0031] As a further aspect of the present invention, the method for performing the association determination process is as follows:

[0032] Clustering is performed on individual deviation ratios to extract potential individual limitation features;

[0033] Analyze the group deviation ratio using scenario-related data to determine if there are any limitations in group behavior.

[0034] Individual limitations are verified based on potential common limitations to determine whether there are limitations in individual behavior.

[0035] As a further aspect of the present invention: the method for extracting the misjudgment time window is as follows:

[0036] The intent feature dataset is filtered, and a three-dimensional association table is constructed based on the filtering results;

[0037] Time windows are extracted from the three-dimensional association table to obtain the misjudgment time windows;

[0038] The validity of the misjudgment time window is verified, and the misjudgment time windows that meet the validity requirements are selected.

[0039] As a further aspect of the present invention: the preferred control window is determined as follows:

[0040] Obtain the frequency of repetition of the same misjudgment time window in the associated chain, as well as the instruction response delay between the start time of the misjudgment time window and the key instruction issuance time in the simulation phase;

[0041] Based on the response capability of the vibration control module of the flight simulator, controllable evaluation conditions are formulated.

[0042] If the repetition frequency and command response delay meet the controllable evaluation conditions, then the effective misjudgment window is marked as the preferred control window.

[0043] As a further aspect of the present invention, the method for obtaining the association chain is as follows:

[0044] Obtain the same feature ID and operation intent type from the three-dimensional association table and combine them, then count the frequency of occurrence of the combination in each simulation stage and the total frequency of occurrence;

[0045] Calculate the proportion of the frequency of each combination in each simulation phase to the total frequency of each combination, and obtain the frequency proportion of each combination in each simulation phase;

[0046] Obtain the duration of each simulation phase and the total flight time, calculate the proportion of each simulation phase in the total flight time, and obtain the phase duration proportion of each simulation phase.

[0047] If the frequency of combinations in the simulation phase is higher than the duration of the phase, then the corresponding simulation phase is marked as the main associated phase of the combination.

[0048] If a combination has been identified as a potential personality limitation trait, then a linking chain is formed by binding the pilot number, trait ID, operational intent type, and main association stage.

[0049] The beneficial effects of this invention are:

[0050] (1) Collect pitch vibration signals from the control stick and extract effective vibration segments. At the same time, combine scene state parameters and control force parameters to complete the operation intention association marking and construct an intention feature dataset. This is beneficial to improve the effectiveness of pitch vibration signals, reduce the impact of equipment mechanical errors and environmental interference on feature extraction, and make the operation intention marking fit the actual flight operation scenario, providing a data foundation for subsequent intention analysis and control optimization.

[0051] (2) Based on the intention feature dataset, establish the operational baseline for each pilot, identify the accidental characteristics of behavior through deviation analysis, and group the statistical features-intention co-occurrence patterns by pilot number and screen the pseudo-correlated feature-intention pairs. This helps to distinguish the fluctuation of accidental features under the same operational intention from the routine operational features, reduce the mis-correlation of features-intention caused by instantaneous interference, make the identification of pseudo-correlated features more targeted, and provide a basis for positioning and control problems.

[0052] (3) Calculate the individual deviation ratio and the group deviation ratio according to the pilot number, extract potential individual limitation features through clustering, determine the group behavior limitation through dynamic time warping algorithm, and then associate pseudo-correlation feature-intent pairs at different stages and extract the verification misjudgment time window. This is helpful to determine the limitation type of pseudo-correlation feature-intent pairs, and the extraction of misjudgment time window can focus on the key time interval that needs to be optimized, providing optimization targets for subsequent differentiated control. Attached Figure Description

[0053] The invention will now be further described with reference to the accompanying drawings.

[0054] Figure 1 This is a block diagram of an adaptive control system for pitch vibration of an aircraft according to the present invention;

[0055] Figure 2 This is a flowchart for determining the behavioral limitations in a group setting in this invention;

[0056] Figure 3 This is a flowchart of an adaptive control method for pitch vibration of a flight simulator according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1

[0059] Please see Figure 1 As shown, the present invention is an adaptive control system for pitch vibration of an aircraft, comprising the following modules:

[0060] Feature acquisition module: used to acquire input features of joystick pitch vibration, perform operation intention association labeling on the input features, and construct an intention feature dataset after association labeling.

[0061] The method for acquiring the input characteristics of the joystick pitch vibration is as follows:

[0062] Preferably, the real-time vibration signal of the control stick in the pitch direction is collected based on the three-axis accelerometer and angular displacement sensor deployed in the flight simulator;

[0063] It should be noted that the pitch direction is the direction of rotation up and down along the vertical axis of the flight simulator. The pitch vibration of the flight simulator corresponds to the typical operating conditions of aviation scenarios, including climb vibration during takeoff, turbulence vibration during cruise, dive correction vibration during landing, and sudden pitch vibration in the scenario;

[0064] Real-time vibration signals are analyzed and processed using a wavelet threshold denoising algorithm to extract effective vibration segments.

[0065] Those skilled in the art will understand that the vibration analysis and processing is performed as follows: 100 sets of unloaded vibration data marked by technicians as having no operational intent are acquired; the wavelet threshold denoising algorithm trains a noise model of the flight control stick vibration by loading pre-trained unloaded vibration data; the noise model processes the acquired real-time vibration signal segment by segment, identifies and filters mechanical errors of the equipment and environmental interference, and segments the effective feature segments corresponding to the true intent by using a preset continuity threshold for the vibration signal.

[0066] The continuity threshold of the preset vibration signal is determined based on the statistical analysis of 500 sets of effective vibration segments with known true operational intentions. Specifically, the amplitude variation coefficient corresponding to three consecutive effective vibration segments is ≤15%, and the variation coefficient is the ratio of the standard deviation to the mean. The continuity threshold is used to ensure that the segmented segments are all driven by coherent operational intentions and to exclude discrete data segments caused by instantaneous noise.

[0067] Extract vibration frequency, amplitude, vibration duration, and instantaneous impact intensity from each effective feature segment of the real-time vibration signal;

[0068] Vibration frequency, amplitude, duration of vibration, and instantaneous impact intensity are used as input features;

[0069] The process of associating input features with operational intent and constructing the associated intent feature dataset is as follows:

[0070] S101. Transform the input features into an input feature vector, and determine the classification dimension of the operation intent; obtain historical input feature vectors and combine them with the classification dimension to construct an intent feature vector;

[0071] Collect the direction vector of the axial force applied by the pilot to the control stick and the direction of vibration;

[0072] The direction vector and input features are integrated to construct the input feature vector;

[0073] Preferably, based on the behavioral logic of flight simulation operations, operational intentions are divided into active control intentions, disturbance response intentions, and misoperation intentions;

[0074] It should be noted that active control intention refers to the pitch operation initiated by the pilot to achieve a preset flight attitude, such as adjusting the climb angle or maintaining the cruise altitude.

[0075] Disturbance response intent: Passive corrective actions taken by the pilot in response to external disturbances generated by the simulator, such as simulated strong airflow or engine thrust fluctuations;

[0076] Intended action: Unintended pitch vibration caused by mechanical delay of the control stick or pilot error;

[0077] Extract the input feature vector corresponding to each operation intention from the historical operation logs of the flight simulator, and use it as the intention feature vector;

[0078] S102. Collect and verify supplementary data, and perform matching and labeling processing on the input feature vector based on the historical intent feature vector to obtain the initially labeled operation intent;

[0079] The verification supplementary data includes: scene state parameters output by the flight simulator, and operating force parameters output by the joystick;

[0080] For example, scenario state parameters include the current simulation phase (e.g., takeoff, landing, cruise), preset disturbance type (e.g., airflow disturbance, mechanical failure), and disturbance intensity level;

[0081] Operating force parameters include: axial force amplitude and force duration;

[0082] The method for matching and labeling the input feature vector based on historical intent feature vectors is as follows:

[0083] A vector similarity matching algorithm is used to match and label the input feature vector with the historical intent feature vector to obtain the initially labeled operation intent;

[0084] Those skilled in the art will understand that the method of matching and labeling the input feature vector with the historical intent feature vector using a vector similarity matching algorithm is as follows:

[0085] Calculate the cosine similarity between the input feature vector to be labeled and the historical intent feature vectors: Let A be the set of historical active control intent feature vectors, B be the disturbance response intent, and C be the misoperation intent. Calculate the average cosine similarity between the vector to be labeled and all vectors in A, B, and C respectively to obtain the similarity value. , , ;

[0086] Preliminary marking: If is the maximum value and A similarity score ≥0.7 (0.7 is the preset similarity threshold) is initially marked as an intention to actively regulate.

[0087] like is the maximum value and ≥0.65, initially marked as a disturbance response intention; if is the maximum value and ≥0.6, initially marked as an intentional erroneous operation;

[0088] The preset similarity thresholds (0.7, 0.65, 0.6) were determined by ROC curve analysis of 500 historical labeled samples, and the results of the ROC curve analysis were manually screened.

[0089] S103. Combine the initially labeled operation intention with the verification supplementary data to perform rule verification and correction, obtain the operation intention label, and combine the input feature vector to realize the association labeling process;

[0090] For example, the rule verification and correction method is as follows: if it is initially marked as an active control intention, the scene state parameters must be undisturbed and the operation force amplitude must be ≥5N; otherwise, it is corrected to a disturbance response intention (if the scene contains disturbances) or a misoperation intention (if the operation force is not up to standard).

[0091] If it is initially marked as a disturbance response intention, it must meet the following conditions: the scene state parameters contain disturbances and the amplitude of the operating force is positively correlated with the disturbance intensity level (for example, the force amplitude increases by 1 to 2 N for every 1 level increase in the level); otherwise, it should be corrected to an active control intention.

[0092] The positive correlation between the amplitude of the operating force and the level of disturbance intensity was verified by linear regression fitting of operating force data under 200 disturbance scenarios. The obtained linear correlation coefficient R0 was... 2 ≥0.85, ensuring that the correlation between grade improvement and force amplitude increase is statistically significant, and that technicians can reproduce this relationship through the same data fitting process;

[0093] If the initial labeling is a misoperation intention, the following conditions must be met: the amplitude of the operating force ≤ 3N and the angle between the direction vector and the vibration direction > 90° (i.e., it is in a non-target direction); otherwise, it will be relabeled based on the second largest similarity value.

[0094] The input feature vector, orientation vector, scene state parameters, operation force parameters, and operation intent labels are associated and combined to construct an intent feature dataset.

[0095] Intent prediction module: Extracts incidental behavioral features from the intent feature dataset, performs correlation screening between incidental behavioral features and operational intent to obtain pseudo-correlation feature-intent pairs;

[0096] The method for extracting incidental behavioral features from the intent feature dataset is as follows:

[0097] It should be noted that random behavioral characteristics refer to irregular fluctuations in characteristics that occur under the same operational intent. The main difference between random behavioral characteristics and regular operational characteristics is that regular operational characteristics have a clear sequential dependency, while random behavioral characteristics, due to instantaneous interference, do not have a stable correlation between preceding and following characteristics.

[0098] S201. Establish an operational baseline for each pilot based on the intent feature dataset;

[0099] Preferably, the method for establishing the operational baseline is as follows: for each type of operational intent, extract all valid segment input feature vectors of a single pilot as operational samples; set a sliding time window (window length = 0.5s, based on a sampling frequency of f = 20Hz, containing 10 sampling points), and calculate the mean and standard deviation of each input feature (vibration frequency, amplitude, etc.) corresponding to the operational samples within the sliding time window, which are used as the baseline values ​​of each input feature under the corresponding operational intent of the pilot (the mean is the baseline center value, and the standard deviation is the baseline fluctuation range).

[0100] S202. Real-time acquisition of the pilot's input feature vector, and deviation analysis of the input feature vector from the operating baseline to obtain accidental judgment features;

[0101] The preferred method for performing flow deflection analysis is as follows:

[0102] The pilot's input feature vector is collected in real time, and the input feature vector is matched with the operation intention type according to the matching and labeling method based on historical intention feature vector in S102.

[0103] The absolute deviation ratio between each matched input feature and the baseline value of each input feature corresponding to each operation intent type is calculated to obtain the deviation degree of a single input feature.

[0104] It should be noted that the deviation is the absolute value of the difference between the real-time feature value and the baseline value, which is then calculated by comparing the absolute value of the difference with the baseline value.

[0105] Calculate the mean of the deviation of each individual input feature across all types of input features, and use it as the feature deviation.

[0106] Obtain the number of samples with continuous deviations, and the number of regression samples whose features regress to the baseline ±0.1 after the deviations are generated;

[0107] The feature deviation, the number of consecutively deviating samples, and the number of regression samples are used as randomness determination features;

[0108] S203. Establish random feature determination rules. If a random feature satisfies the random determination rules, then mark the current input feature vector as a random behavioral feature.

[0109] For example, the way to establish the random feature judgment rule is as follows: Judgment condition 1: Deviation D > 0.2 (exceeding the normal fluctuation range of individual stable operation).

[0110] Judgment condition two: The number of consecutive deviation samples N≤5 (sampled at 20Hz, duration ≤0.25 seconds, shorter than the minimum duration of intentional operation adjustment);

[0111] Judgment condition three: The number of regression samples is 3, that is, within 3 samples after deviation (≤0.15 seconds), the feature value regresses to the range of f_base±0.1 (that is, fast regression to the benchmark, no continuous deviation trend, excluding active adjustments with clear objectives);

[0112] If all three judgment conditions are met, then the current input feature vector is determined to be a random feature.

[0113] Conversely, no action is taken;

[0114] Those skilled in the art will understand that the deviation threshold (0.2) and regression range (±0.1) are derived from statistics of 2000 samples marked as stable operations (boundary values ​​of the 95% confidence interval), and technicians can obtain parameters adapted to a specific simulator through the same process;

[0115] The process of filtering out false correlations between incidental behavioral features and operational intentions to obtain pseudo-correlated feature-intention pairs is as follows:

[0116] The samples in the intent feature dataset are randomly divided into K groups according to the pilot ID;

[0117] Preferably, K=3, and each group contains ≥30 pilots to ensure a balanced sample size. Each group is denoted as Group1, Group2, and Group3.

[0118] For each incidental behavioral feature, the co-occurrence pattern of the feature and intention is statistically analyzed in K groups, that is, the proportion of the incidental behavioral feature in the three operational intentions of active control, disturbance response, and misoperation.

[0119] For example, in Group 1, the proportion of features in the occurrence of active control intentions is 30%, in disturbance response is 50%, and in misoperation is 20%.

[0120] The similarity of K co-occurring patterns is calculated using the cosine similarity algorithm, and highly consistent pairs are extracted based on the similarity.

[0121] The high-consistency groups are processed for pseudo-correlation detection to obtain pseudo-correlation feature-intent pairs;

[0122] For example, the method for extracting highly consistent pairs based on similarity and handling pseudo-correlation of highly consistent pairs is as follows: cosine similarity is used to calculate the co-occurrence pattern similarity between Group1 and Group2, and between Group1 and Group3, respectively. The similarity value ranges from [0,1], and the closer it is to 1, the more consistent the patterns of the two groups are; the cosine similarity of the co-occurrence patterns of the two groups of samples is ≥0.8, and the statistical range of the pairs is Group1-Group2, Group1-Group3, and Group2-Group3 (a total of 3 pairs); if the number of highly consistent pairs in the 3 pairs is ≤1, it is judged as a pseudo-correlation feature-intent pair; if the number of highly consistent pairs is ≥2, it is judged as a stable association and is not included in the pseudo-correlation range.

[0123] Example 2

[0124] Please see Figure 1 As shown, the present invention is an adaptive control system for pitch vibration of an aircraft, which also includes the following modules:

[0125] The adaptation analysis module analyzes the pseudo-relevant feature-intent pairs through behavioral limitation analysis to determine whether there are behavioral limitations; if so, it performs misjudgment window identification processing on the pseudo-relevant feature-intent pairs to obtain the misjudgment time window.

[0126] Among them, the method for determining whether there is behavioral limitation in spurious relevance feature-intent pairs through behavioral limitation analysis is as follows:

[0127] Collect the frequency of occurrence of pseudo-correlation features—intent pairs—for each pilot according to their identification number, and the total frequency of pilot operations.

[0128] It should be noted that the total operation frequency is the total number of valid operation samples for each pilot in the intention feature dataset corresponding to the pseudo-correlation feature-intention pair under the operation intention (i.e. the total number of all valid vibration segments of the pilot under that intention).

[0129] The individual deviation ratio is obtained by calculating the proportion of the occurrence frequency of each pilot's pseudo-correlation feature-intent pair to the total operation frequency.

[0130] Where, individual deviation = (the pilot's pseudo-correlation feature - the frequency of occurrence of the intention pair) / (the pilot's total operation frequency under the corresponding operation intention;

[0131] The mean of the individual deviation ratios of all pilots is calculated to obtain the group deviation ratio;

[0132] The association determination is performed based on the individual deviation ratio and the group deviation ratio to obtain the result of whether the behavior is restrictive;

[0133] S301. Cluster the individual deviation ratios to extract potential individual limitation features;

[0134] Preferably, the clustering method for individual deviation ratios is as follows: The unsupervised clustering algorithm (DBSCAN) is used to perform cluster analysis on the individual deviation ratios of all pilots.

[0135] Using individual deviation ratio as the clustering feature, clusters are automatically divided based on data density. The clustering algorithm uses neighborhood radius and minimum sample number parameters (adaptively determined based on the distribution of individual deviation ratios of 50 pilots to ensure high data similarity within clusters and significant differences between clusters).

[0136] If a pilot's individual deviation ratio forms an isolated cluster, and the isolated cluster appears repeatedly under the same operational intent in different simulation scenarios (e.g., isolated clusters in both takeoff and cruise scenarios), then the pilot's pseudo-correlation feature-intent pair is marked as a potential personality limitation feature.

[0137] S302. Conduct scenario-related analysis on the group deviation ratio to determine whether there are any limitations in group behavior;

[0138] Preferably, the group deviation ratio and the disturbance intensity level in the scene state parameters are obtained for N monitoring periods;

[0139] Construct time series curves for population deviation ratio and disturbance intensity level, and use dynamic time warping algorithm to measure the similarity between the two time series curves;

[0140] like Figure 2 As shown, if the similarity is higher than or equal to the preset similarity threshold, it is determined that there are behavioral limitations in the group scenario.

[0141] If the similarity is lower than the preset similarity threshold, the change in similarity will be continuously monitored.

[0142] S303. Verify individual limitations based on potential common limiting characteristics to determine whether there are limitations in individual behavior;

[0143] Preferably, the time window of occurrence of potential common limiting features in each pilot's operation record is extracted (determined based on the sampling timestamp of the feature acquisition module), and the operation action sequence data of the pilot within the corresponding time window is retrieved simultaneously (recorded by the six-axis motion capture sensor built into the control stick, including the push stick force change curve, the control stick pitch angle adjustment trajectory, and the force application point offset); using the operation action sequences of 50 pilots in the same common operation scenario (the scenario associated with potential common limiting features) as samples, a hierarchical clustering algorithm is used to generate a group standard operation action sequence cluster (the action sequence corresponding to the cluster center is taken as the standard to ensure that the cluster contains ≥80% of the pilots' action sequences).

[0144] The pilot's action sequence is compared with the standard action sequence cluster of the group. If the pilot's action sequence forms an individual-unique sub-cluster independent of the standard cluster in all feature dimensions (such as force, angle, and point of application) (i.e., the Euclidean distance between the sub-cluster and the standard cluster is greater than the maximum distance within the standard cluster), and the time window of this sub-cluster coincides with the time window of potential common limiting features, it indicates that the individual has unique operational deviations (such as excessive stick push force) in the common scenario of the group (such as the strong airflow disturbance stage), so it is determined that there are individual behavioral limitations; if the individual action sequence within the overlapping window is consistent with the group standard cluster, then individual limitations are excluded.

[0145] It is understandable that determining the behavioral limitations in group settings serves the following purpose:

[0146] Function 1: Differentiating the root causes of problems and reducing deviations in optimization direction; identifying the existence of behavioral limitations in a group scenario can clarify the common attributes of pseudo-related feature-intent pairs, distinguishing between common group problems and individual operational problems. By identifying the root causes of problems (such as unreasonable scene parameter settings or common defects in equipment hardware), this provides a basis for subsequent targeted optimization and reduces the waste of resources from meaningless individual differences in adjustments.

[0147] Secondly, it supports the extraction of common misjudgment windows for groups, improving optimization efficiency; determining the existence of behavioral limitations in group scenarios is a prerequisite for subsequently extracting the time window for group co-occurrence misjudgment. The misjudgment window of group co-occurrence has the characteristics of wide coverage and high optimization benefits. The subsequent compensation evaluation module can formulate a unified common optimization strategy for such windows to improve optimization efficiency;

[0148] Thirdly, it helps ensure the consistency of flight simulation training scenarios and the effectiveness of training; identifying group behavioral limitations in scenarios can avoid training deviations caused by common problems. If group behavioral limitations are not identified, multiple pilots will have consistent misjudgments of operational intentions in the same simulated scenario, causing the simulation training feedback to become disconnected from the logic of real flight operations, affecting the pilots' understanding of standardized operations. By identifying and subsequently optimizing group limitations, the actual effect of flight simulation training can be improved.

[0149] The method for identifying and extracting misjudgment time windows by associating false correlation features of pilots at different stages is as follows:

[0150] S311. Filter the intent feature dataset and construct a three-dimensional association table based on the filtering results;

[0151] Preferably, the pseudo-correlation feature-intent pairs of the pilot in each stage are filtered from the intent feature dataset by using the simulation stage labels (including takeoff stage, cruise stage, landing stage, and emergency scenario stage) output by the flight simulator scenario control module as the dimension and the pilot number + operation intent type.

[0152] The filtering results are integrated into a three-dimensional association table, with fields including: pilot number, operation intention type, stage label, feature ID, feature first frame timestamp, and feature last frame timestamp.

[0153] The feature last frame timestamp is calculated based on a 20Hz sampling frequency (i.e., signal sampling frequency) and feature duration: last frame timestamp = first frame timestamp + feature duration × 1000ms;

[0154] S312. Extract time windows from the three-dimensional association table to obtain the misjudgment time window;

[0155] Preferably, the feature time window is initially determined by taking the timestamp of the first frame as the starting point and the timestamp of the last frame as the ending point;

[0156] Retrieve the true value log of the flight simulator's operation intent (the true intent time interval jointly annotated by 3 or more annotators based on operation video and equipment data, including clear intent type labels). If the feature-marked intent (i.e. intent type label) is inconsistent with the true intent, the feature time window is directly used as the confirmation misjudgment time window.

[0157] If the intention of feature labeling is different from the truth value... Figure 1 If the feature is a false correlation (i.e., the correlation is unstable), then the overlap between the feature time window and the true value intention interval is taken as the latent misjudgment time window.

[0158] S313. Verify the validity of the misjudgment time window and filter out the misjudgment time windows that meet the validity requirements.

[0159] Preferably, for the extracted misjudgment time window, if the corresponding feature is a personal limitation feature, verify the degree of overlap between the window and the time window of the pilot's individual unique sub-cluster action sequence.

[0160] It is understandable that the overlap ratio is the ratio of the window intersection duration to the window union duration;

[0161] If the corresponding feature is a common limitation feature, verify the overlap between the verification window and the time window of abnormal fluctuation of equipment parameters; windows with an overlap of ≥90% are included in the effective misjudgment time window library and stored with the pilot number feature ID and stage label as unique identifiers;

[0162] Compensation and evaluation module: Performs controllable evaluation processing on the misjudgment time window, determines the optimal control window, optimizes pitch vibration control on the pseudo-correlation feature-intent pair based on the optimal control window, collects matching data of optimized pitch vibration and operation intent, and evaluates it;

[0163] The method for controllingly evaluating and determining the optimal control window for misjudgment time windows is as follows:

[0164] Obtain the same feature ID and operation intent type from the three-dimensional association table and combine them, then count the frequency of occurrence of the combination in each simulation stage and the total frequency of occurrence;

[0165] Calculate the proportion of the frequency of each combination in each simulation phase to the total frequency of each combination, and obtain the frequency proportion of each combination in each simulation phase;

[0166] It should be noted that the simulation phase refers to the simulation phase in the feature acquisition module (e.g., takeoff, landing, cruise).

[0167] Obtain the duration of each simulation phase and the total flight time, calculate the proportion of each simulation phase in the total flight time, and obtain the phase duration proportion of each simulation phase.

[0168] If the frequency of combinations in the simulation phase is higher than the duration of the phase, then the corresponding simulation phase is marked as the main associated phase of the combination.

[0169] If a combination has been identified as a potential personality limitation trait, then a linking chain is formed by binding the pilot number, trait ID, operational intent type, and main association stage.

[0170] Obtain the frequency of repetition of the same misjudgment time window in the associated chain, as well as the instruction response delay between the start time of the misjudgment time window and the key instruction issuance time in the simulation phase;

[0171] Based on the response capability of the vibration control module of the flight simulator, controllable evaluation conditions are formulated.

[0172] If the repetition frequency and command response delay meet the controllable evaluation conditions, then the effective misjudgment window is marked as the preferred control window;

[0173] It is understandable that the purpose of obtaining the preferred control window is:

[0174] Function 1: Focus on effective optimization targets; Filter out windows that meet controllable conditions (repetition frequency meets the standard, response delay matches module capabilities, etc.) from the misjudgment time window, reduce ineffective adjustments to uncontrollable and occasional misjudgment windows, and allow vibration control optimization to focus on high-value and improveable key intervals;

[0175] Secondly, it ensures that the optimization strategy can be implemented; it facilitates the adaptation of the optimal control window to the response capability of the flight simulator vibration control module (such as the window duration being within the adjustable range of the module), providing an executable operation carrier for subsequent differentiated adjustments, and avoiding optimization failure due to window parameters exceeding the module's capabilities;

[0176] For example, the controllable evaluation conditions are set as follows: the repetition frequency of the effective misjudgment window reaches the minimum threshold of the corresponding limitation type, and the instruction response delay between its start time and the key instruction issuance time in the simulation stage is ≤ the maximum responsive delay of the module. At the same time, the window duration is within the adjustable range of the module. When all three conditions are met, the repetition frequency and instruction response delay are considered to meet the controllable evaluation conditions.

[0177] The minimum threshold for each limitation type is determined by historical operational data. Individual limitations (specific to a single pilot) are set based on historical data from the pilot's last 10 identical simulation training sessions in the same main correlation phase, with a minimum threshold of ≥3 occurrences (excluding occasional false positives). Common limitations (co-occurring among multiple pilots) are set based on historical data from ≥5 pilots in the same main correlation phase, with a minimum threshold of ≥10 occurrences (ensuring stable false positives for group co-occurrence).

[0178] The maximum response delay of the module is determined by combining the hardware performance of the vibration control module of the flight simulator (such as the response speed of the drive motor and the signal transmission delay) through hardware parameter testing and calibration. It is the longest allowable time for the module to stably execute vibration control commands (the example value is 50ms, and the specific value needs to be matched with the actual performance of the equipment).

[0179] Adjustable range of the module; determined by testing the adjustment effect of different duration windows based on the signal acquisition accuracy (e.g., 20Hz sampling frequency) and control command execution efficiency of the vibration control module, and must be within the duration range in which the module can output vibration adjustment (exemplary range is 10-100ms, ensuring that effective adjustment can be completed within the window).

[0180] Based on the pseudo-correlation feature-intent pairs within the preferred control window, a parameter adaptive adjustment strategy algorithm is constructed to optimize pitch vibration control.

[0181] It is understandable that the method for constructing a parameter adaptive adjustment strategy to optimize pitch vibration control is as follows:

[0182] Differentiated adjustment parameters are applied to address the limitations of pseudo-correlation features-intent pairs within the preferred control window. If a unique limitation feature is identified, the pilot's individual sub-cluster action sequence data (such as stick force fluctuation curve and stick angle adjustment trajectory) is called up. The dynamic changes in vibration frequency and amplitude are collected in real time within the window. The stick feedback damping coefficient is adaptively adjusted based on the principle of reverse cancellation of action fluctuation trends.

[0183] For example, the damping coefficient can be adjusted as follows: for every 0.1 N / ms increase in the rate of force change, the operating lever feedback damping coefficient increases by 0.05 (initial damping coefficient = 0.3). The maximum value after adjustment shall not exceed 0.8. (To avoid excessive damping affecting operation).

[0184] If common limiting characteristics are identified, abnormal fluctuation data of associated equipment parameters (such as fluctuations in the mechanical damping of the control lever and drift of sensor signals) are used. Within the window, the Kalman filter algorithm is enabled to compensate for the vibration signal in real time. The filter coefficient is dynamically updated according to the amplitude of abnormal fluctuations in the equipment (such as increasing the filter strength when there are high-frequency fluctuations and decreasing the filter strength when there are low-frequency fluctuations). Through real-time adaptive adjustment of parameters, vibration deviations caused by pseudo-correlation features are suppressed, thereby improving the matching accuracy between pitch vibration and operating intention.

[0185] The criteria for determining high-frequency fluctuations are as follows: vibration frequency ≥ 10Hz (based on vibration frequency statistics at each stage of the flight simulator, with high-frequency vibration accounting for ≤ 15% during the cruise phase and ≤ 30% during sudden scenarios); filter strength adjustment: during high-frequency fluctuations (vibration frequency ≥ 10Hz), the noise covariance Q of the Kalman filter process increases by 20% (Q initial value = 0.01); during low-frequency fluctuations (vibration frequency < 10Hz), Q remains at its initial value; the observed noise covariance R = 0.02 (determined based on measured sensor noise values);

[0186] The method for collecting and evaluating the matching data of optimized pitch vibration and operational intent is as follows:

[0187] Preferably, vibration data is collected and optimized under the same scenario and the same pilot operation.

[0188] The vibration data includes: optimized vibration frequency, amplitude, duration and operation intention label, and an optimized matching dataset is constructed.

[0189] The matching dataset is evaluated by feature extraction to obtain the accuracy improvement and vibration reduction.

[0190] The method for performing matching feature extraction is as follows:

[0191] The difference between the false positive rate (i.e., the ratio of false positives to total matches) of the optimized pseudo-relevant feature-intent pairs in the matching dataset and the false positive rate before optimization is calculated. The difference is then divided by the false positive rate before optimization to obtain the accuracy improvement.

[0192] Calculate the difference between the mean of the optimized feature vibration amplitude in the matched dataset and the value before optimization, and divide it by the mean amplitude before optimization to obtain the vibration improvement degree.

[0193] The improvement in accuracy and the improvement in vibration are used as evaluation indicators. If the improvement in accuracy is ≥30% and the improvement in vibration is ≥25% at the same time, the optimization is deemed effective.

[0194] Otherwise, based on the evaluation results, the evaluation parameters of the preferred control window are readjusted (e.g., the window duration is widened to 80ms), and pitch vibration control optimization is repeated.

[0195] Example 3

[0196] Please see Figure 3 As shown, the present invention is an adaptive control method for pitch vibration of an aircraft, comprising the following steps:

[0197] Step 1: Collect the input features of the joystick pitch vibration, perform operation intention association labeling on the input features, and construct the intention feature dataset after association labeling.

[0198] Step 2: Extract incidental behavioral features from the intent feature dataset, and perform correlation screening between incidental behavioral features and operational intent to obtain pseudo-correlation feature-intent pairs;

[0199] Step 3: Analyze the spurious relevance features-intent pairs using behavioral limitation analysis to determine if there are behavioral limitations; if so, perform misjudgment window identification processing on the spurious relevance features-intent pairs to obtain the misjudgment time window.

[0200] Step 4: Perform controllable evaluation of the misjudgment time window, determine the optimal control window, optimize pitch vibration control based on the optimal control window for the pseudo-correlation feature-intent pair, collect matching data of optimized pitch vibration and operation intent, and evaluate it.

[0201] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An adaptive control system for pitch vibration of an aircraft, characterized in that: Includes the following modules: Feature acquisition module: used to acquire input features of joystick pitch vibration, perform operation intention association labeling on the input features, and construct an intention feature dataset after association labeling. Intent prediction module: Extracts incidental behavioral features from the intent feature dataset, performs correlation screening between incidental behavioral features and operational intent to obtain pseudo-correlation feature-intent pairs; The adaptation analysis module analyzes the pseudo-relevant feature-intent pairs through behavioral limitation analysis to determine whether there are behavioral limitations; if so, it performs misjudgment window identification processing on the pseudo-relevant feature-intent pairs to obtain the misjudgment time window. The method of behavioral limitation analysis is as follows: Collect the frequency of occurrence of pseudo-correlation features—intent pairs—for each pilot according to their identification number, and the total frequency of pilot operations. The individual deviation ratio is obtained by calculating the proportion of the occurrence frequency of each pilot's pseudo-correlation feature-intent pair to the total operation frequency. The mean of the individual deviation ratios of all pilots is calculated to obtain the group deviation ratio; The association determination is performed based on the individual deviation ratio and the group deviation ratio to obtain the result of whether the behavior is restrictive; Compensation evaluation module: Performs controllable evaluation processing on the misjudgment time window, determines the optimal control window, optimizes pitch vibration control based on the optimal control window for pseudo-correlation feature-intent pair, collects matching data of optimized pitch vibration and operation intent, and evaluates it.

2. The adaptive control system for aircraft pitch vibration according to claim 1, characterized in that: The process of associating the operation intent with the tagging process is as follows: The input features are transformed into input feature vectors, and the classification dimension of the operation intent is determined. Historical input feature vectors are obtained and combined with the classification dimension to construct the intent feature vector. Collect and verify supplementary data, and perform matching and labeling processing on the input feature vector based on historical intent feature vectors to obtain the initially labeled operation intent; The operation intent initially labeled is then combined with supplementary verification data for rule verification and correction to obtain the operation intent label. This label is then combined with the input feature vector to achieve associated labeling processing.

3. The adaptive control system for aircraft pitch vibration according to claim 2, characterized in that: The matching tag processing is performed as follows: A vector similarity matching algorithm is used to match and label the input feature vector with the historical intent feature vector to obtain the initially labeled operation intent.

4. The adaptive control system for aircraft pitch vibration according to claim 1, characterized in that: The method for performing the aforementioned correlation misjudgment screening is as follows: Extract samples from the intent feature dataset and randomly divide them into K groups according to pilot ID; For each incidental behavioral feature, feature-intent co-occurrence patterns are statistically analyzed in K groups. The similarity of co-occurrence patterns in K groups is calculated using the cosine similarity algorithm, and highly consistent pairs are extracted based on the similarity. The high consistency pairs are processed for pseudo-correlation to obtain pseudo-correlation feature-intent pairs.

5. The adaptive control system for aircraft pitch vibration according to claim 4, characterized in that: The method for extracting the incidental features of the behavior is as follows: Establish an operational baseline for each pilot based on the intent feature dataset; The pilot's input feature vector is collected in real time, and the deviation between the input feature vector and the operating baseline is analyzed to obtain the random judgment feature. Establish a random feature determination rule. If a random feature satisfies the random feature determination rule, then mark the current input feature vector as a random behavioral feature.

6. The adaptive control system for aircraft pitch vibration according to claim 1, characterized in that: The method for performing the aforementioned association determination process is as follows: Clustering is performed on individual deviation ratios to extract potential individual limitation features; Analyze the group deviation ratio using scenario-related data to determine if there are any limitations in group behavior. Verify individual limitations based on the context in which group behavior is limited, and determine whether there are limitations in individual behavior.

7. The adaptive control system for aircraft pitch vibration according to claim 1, characterized in that: The method for extracting the misjudgment time window is as follows: The intent feature dataset is filtered, and a three-dimensional association table is constructed based on the filtering results; Time windows are extracted from the three-dimensional association table to obtain the misjudgment time windows; The validity of the misjudgment time window is verified, and misjudgment time windows that meet the validity requirements are selected. The filtering method is as follows: Using the simulation stage labels output by the flight simulator scenario control module as the dimension, and according to the pilot number and operation intention type, pseudo-correlation feature-intention pairs of the pilot in each stage are filtered from the intention feature dataset, and the filtering results are integrated to construct a three-dimensional association table.

8. The adaptive control system for aircraft pitch vibration according to claim 1, characterized in that: The preferred control window is determined as follows: Obtain the same feature ID and operation intent type from the three-dimensional association table and combine them, then count the frequency of occurrence of the combination in each simulation stage and the total frequency of occurrence; Calculate the proportion of the frequency of each combination in each simulation phase to the total frequency of each combination, and obtain the frequency proportion of each combination in each simulation phase; Obtain the duration of each simulation phase and the total flight time, calculate the proportion of each simulation phase in the total flight time, and obtain the phase duration proportion of each simulation phase. If the frequency of combinations in the simulation phase is higher than the duration of the phase, then the corresponding simulation phase is marked as the main associated phase of the combination. If a combination has been identified as a potential personality limitation trait, then a linking chain is formed by binding the pilot number, trait ID, operational intent type, and main association stage. Obtain the frequency of repetition of the same misjudgment time window in the associated chain, as well as the instruction response delay between the start time of the misjudgment time window and the key instruction issuance time in the simulation phase; Based on the response capability of the vibration control module of the flight simulator, controllable evaluation conditions are formulated. If the repetition frequency and command response delay meet the controllable evaluation conditions, then the effective misjudgment window is marked as the preferred control window.

Images