A method for automatic recognition of flattening opportunity based on improved random forest model

Abstract

The present application belongs to the technical field of flight training data processing, and particularly relates to a method for automatically identifying a pullback opportunity based on an improved random forest model, which solves the problems of low efficiency of existing pullback opportunity identification relying on manual operation, single feature of traditional models and poor generalization. The present application obtains QAR data of a flight training simulator, performs preprocessing through cleaning and sample construction; constructs a feature set containing operation response dynamic coupling features and dynamic operation frequency factors; adopts an improved random forest (weighted information gain split criterion + local entropy difference adaptive pruning) training model, combines with adjacent interval fusion, short-time false detection filtering and start time anchoring post-processing to output the pullback opportunity. The present application improves the recognition accuracy, can be adapted to different models, and has no manual intervention throughout, greatly improving the batch processing efficiency of QAR data.

Description

Technical Field

[0001] This invention belongs to the field of flight training data processing and leveling machine identification technology, specifically involving an automatic leveling machine identification method based on an improved random forest model. Background Technology

[0002] The leveling maneuver during aircraft landing is a critical stage of flight. The accuracy of leveling directly affects landing safety and stability. For example, leveling too early may cause the aircraft to "drift" over the runway, prolonging the landing distance; while leveling too late may result in a hard landing, increasing the impact load on the landing gear and fuselage. Currently, in flight training, the determination of leveling accuracy mainly relies on manual review of QAR data (Quick Access Recorder, QAR device; the QAR device built into the simulator can collect multiple flight parameters during simulated flight training, including parameters such as aircraft attitude, speed, altitude, weight, front and rear landing gear status, and pilot operation information). This method has the following shortcomings:

[0003] Inefficient: Manually analyzing time-series QAR data segment by segment consumes a lot of time and is difficult to meet the needs of rapid processing of batch training data;

[0004] Highly subjective: The judgment results depend on the flight experience of the evaluators, and different personnel have different judgment criteria for "flattening the start time", resulting in poor consistency of results;

[0005] Single feature: Traditional automatic identification methods rely only on basic parameters such as radio altitude and pitch angle, ignoring the dynamic coupling relationship between "crew control and aircraft response", resulting in low identification accuracy in complex flight scenarios (such as crosswinds and turbulence).

[0006] Poor model generalization: Existing decision tree models are prone to overfitting due to excessively deep branches, and their adaptability to different models and training scenarios is insufficient.

[0007] To address the aforementioned problems, this invention proposes a method for automatic identification based on a leveling process using an improved random forest model. By optimizing data preprocessing, feature construction, and model structure, it overcomes the inefficiency of manual judgment and the limitations of traditional models. Summary of the Invention

[0008] To address the aforementioned problems in existing technologies, namely the low efficiency, high subjectivity, limited feature set, and poor model generalization of aircraft leveling and identification, which rely on manual review of QAR data, this invention proposes a method for automatic leveling and identification based on an improved random forest model. This method includes the following steps:

[0009] Acquire QAR data from the flight training simulator; the QAR data includes radio altitude, pitch angle, rate of pitch change, rate of descent, training tags, and pitch direction control values ​​for the left and right seat stick positions.

[0010] The QAR data is preprocessed and a feature set is constructed. The preprocessing includes data cleaning and sample construction. The sample construction includes merging the pitch direction control values ​​of the left and right side sticks to generate a comprehensive control value.

[0011] The features in the feature set include dynamic coupling features of operation response constructed based on the cumulative control intensity and residuals of the comprehensive control quantity, and dynamic operation frequency factor calculated based on the proportion of samples in the comprehensive control quantity that exceed the dynamic threshold within the sliding window; wherein, the residuals are derived based on the difference between the actual pitch angle difference value and the pitch angle predicted response value.

[0012] The feature set is input into a pre-trained improved random forest model, and the output is an initial flattened label sequence; wherein, the improved random forest model uses weighted information gain as the splitting criterion when splitting decision tree nodes, and the weighted information gain is calculated by weighting the traditional information gain of the feature, the historical weight factor and the dynamic operation frequency factor; the improved random forest model adopts an adaptive pruning strategy during the branch growth stage of the decision tree.

[0013] The initial flattening marker sequence is subjected to neighbor interval fusion, short-term false detection filtering and start time anchoring to obtain the flattening time stamp.

[0014] In some preferred embodiments, the construction of the dynamic coupling feature of the operation response includes:

[0015] Calculate the integral of the overall manipulation amount within the predetermined forward window, and use this integral as the cumulative manipulation intensity;

[0016] The integrated control inputs, combined with a first-order hysteresis system model, yield the pitch angle prediction response value.

[0017] Calculate the residual between the actual pitch angle difference and the predicted pitch angle response value;

[0018] Multiplying the residual by the cumulative manipulation intensity yields the dynamic coupling characteristics of the operation response.

[0019] In some preferred embodiments, the discrete form of the first-order hysteresis system model is as follows:

[0020] ;

[0021] ;

[0022] in, Let be the predicted response value of the pitch angle at time t. Let K be the predicted response value of the pitch angle at time t-1, and K be the gain coefficient. The lag coefficient, The operation response lag time constant, The sampling interval for QAR data. Let t be the total manipulation quantity at time t.

[0023] In some preferred embodiments, the formula for calculating the dynamic operation frequency factor is:

[0024] ;

[0025] ;

[0026] in, For dynamic operation frequency factor, for The total manipulation amount at each time step, where N is the total number of time steps to be summed, accumulated from 1 to N. This is an indicator function; if the condition within the square brackets is true, I = 1; otherwise, I = 0. Here, `median` is the dynamic threshold, `median` is the median function, and `MAD` is the median absolute deviation function. In the time window The overall manipulation amount, where λ is the adjustment factor. Sliding window width.

[0027] In some preferred embodiments, the weighted information gain formula is:

[0028] ;

[0029] in, The weighted information gain of feature A on the sample set S. For the traditional information gain of feature A on sample set S, The historical weighting factor for feature A.

[0030] In some preferred embodiments, the adaptive pruning strategy is based on local entropy difference: when the decrease in the average information entropy of the child nodes generated by node splitting is less than the decrease in the information entropy of the parent node, the current branch expansion is terminated.

[0031] In some preferred embodiments, the specific processes of neighbor interval fusion, short-term false detection filtering, and start time anchoring are as follows:

[0032] The neighboring interval fusion: For consecutive intervals marked as 1 in the initial flattening mark sequence, if the interval between any two segments does not exceed one QAR data sampling period and at least one segment has a duration greater than two QAR data sampling periods, they are merged into one flattening event interval;

[0033] The short-term false detection filtering: For all flattened event intervals after fusion, if there are multiple intervals and their intervals are more than one QAR data sampling period, and the duration of each interval does not exceed two QAR data sampling periods, then all these intervals will be filtered out.

[0034] Starting time anchoring and output: In the flattened event interval after fusion and filtering, the starting timestamp of the first interval is selected as the final flattening output.

[0035] In some preferred embodiments, the data cleaning involves identifying and discarding invalid QAR data with incomplete flight segment information or abnormal parameter jumps, while retaining valid QAR data.

[0036] In some preferred embodiments, the sample construction includes calculating the first-order difference features of pitch angle and descent rate based on the effective data to obtain pitch angle difference features and descent rate difference features, and deleting the first row of null values ​​generated by the difference operation;

[0037] The generated pitch angle difference feature and descent rate difference feature are used to replace the original pitch angle and descent rate parameters, respectively, as dynamic features;

[0038] Based on the dynamic characteristics and the comprehensive manipulation amount, the manually labeled starting time of flattening is located as a key sample, and the first N rows of the key sample are taken as a reference sample reflecting the state before flattening.

[0039] The key samples and the benchmark samples are filtered for duplicate samples through conditional judgment and then added to the feature set.

[0040] In some preferred embodiments, the training label is a manually labeled flattening start time. The labeling rules for manual labeling include: the first timestamp where the pitch angle begins to rise continuously and the descent rate begins to decrease continuously is taken as the flattening start time and labeled as 1, and the remaining timestamps are labeled as 0.

[0041] The beneficial effects of this invention are:

[0042] 1) By capturing the correlation between crew control and aircraft response through "dynamic coupling features of operational response", the recognition accuracy is improved compared with traditional basic features;

[0043] 2) Improve the combination of "weighted information gain + adaptive pruning" in random forest to avoid overfitting, improve adaptability to complex scenarios such as crosswinds and turbulence, and reduce false detection rate;

[0044] 3) Only the parameters of the first-order lag model need to be adjusted to adapt to QAR data of different aircraft types (such as Boeing 737 and Airbus A320), without the need to reconstruct the model structure;

[0045] 4) No manual intervention is required throughout the process, and the efficiency of batch processing QAR data is greatly improved compared to manual review. Attached Figure Description

[0046] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0047] Figure 1 This is a flowchart of the steps of an automatic identification method for leveling based on an improved random forest model according to the present invention. Detailed Implementation

[0048] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0049] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] To more clearly explain the automatic identification method for leveling based on an improved random forest model of the present invention, the following will be combined with... Figure 1 The steps in the embodiments of the present invention will be described in detail below.

[0051] This invention proposes a method for automatic identification during leveling based on an improved random forest model, see [link to relevant documentation]. Figure 1 The method includes the following steps:

[0052] Acquire QAR data from the flight training simulator; the QAR data includes radio altitude, pitch angle, rate of pitch change, rate of descent, training tags, and pitch direction control values ​​for the left and right seat stick positions.

[0053] In this embodiment, the QAR data in the flight training simulator is decoded and parsed to obtain a CSV file;

[0054] Extract the QAR data required for leveling from the CSV file. This data is time-series sampled data (sampling period can be set to 1Hz~10Hz), including radio altitude (ra), pitch angle (pitch), pitch rate of change (pitch_rate), descent rate (descentrate), pitch direction of left side stick position (col_hpos_capt), pitch direction of right side stick position (col_hpos_fo), and training labels. The training labels are manually labeled leveling start times. The manual labeling rules include: the first timestamp where the pitch angle begins to rise continuously and the descent rate begins to decrease continuously is taken as the leveling start time and labeled as 1, and the remaining timestamps are labeled as 0.

[0055] The QAR data is preprocessed and a feature set is constructed. The preprocessing includes data cleaning and sample construction. The sample construction includes merging the pitch direction control values ​​of the left and right side sticks to generate a comprehensive control value.

[0056] In this embodiment, the data cleaning involves identifying and discarding invalid QAR data that has incomplete flight segment information (does not include the entire landing process of "approach-flatten-touch") or has abnormal parameter jumps (e.g., the difference in parameter between adjacent sampling times exceeds 3 times the normal fluctuation range of the parameter, such as the difference in pitch angle between adjacent times > 3°, or the difference in descent rate between adjacent times > 500 ft / min), while retaining valid QAR data;

[0057] The sample construction includes, based on the valid data, calculating the first-order difference of the pitch angle to obtain the pitch angle difference feature (pitch_diff = pitch(t) - pitch(t-1)), calculating the first-order difference of the descent rate to obtain the descent rate difference feature (descentrate_diff = descentrate(t) - descentrate(t-1)), and deleting the first row of null values, i.e., NaN values, generated by the difference operation (because there is no t-1 data at t=0).

[0058] The generated pitch angle difference feature and descent rate difference feature (pitch_diff, descentrate_diff) are used to replace the original pitch angle and descent rate parameters (pitch, descentrate) respectively, as dynamic features that better reflect the changing trend of the parameters and fit the dynamic characteristics of the flattening operation.

[0059] The pitch control values ​​of the left and right side sticks are combined to generate a comprehensive control value hpos_sum (t) = col_hpos_capt (t) + col_hpos_fo (t), which reflects the overall pitch control intention of the unit and avoids the one-sidedness of single-seat control values.

[0060] Based on the dynamic features and the comprehensive manipulation amount, the manually labeled starting time of flattening (corresponding to the row with training label 1) is located as a key sample to reflect the flattening state. The first N rows of the key sample are taken as the benchmark sample to reflect the state before flattening. N is a preset value. In this embodiment, N=6 is preferred. If there are less than 6 rows before the key sample, the first row of valid data is taken.

[0061] After filtering out duplicate samples from the key samples and the benchmark samples through conditional judgment, they are added to the feature set. Specifically, the uniqueness of the timestamp can be used to determine whether the samples are duplicated. If the timestamp of the benchmark sample has been included in the training set, it is skipped; otherwise, it is retained. Key samples have a very low probability of duplication because they are marked with a unique flattened start time. If they are duplicated, they are also filtered out.

[0062] The feature set includes dynamic coupling features for operation response constructed based on the cumulative control intensity and residuals of the integrated control quantity, and a dynamic operation frequency factor calculated based on the proportion of samples where the integrated control quantity exceeds a dynamic threshold within a sliding window; wherein, the residual is derived from the difference between the actual pitch angle difference value and the predicted pitch angle response value; specifically, to improve the accuracy and robustness of leveling operation recognition, a spatiotemporal fusion method of pitch angle difference and control stick integral quantity is used to construct a more physically interpretable composite dynamic feature. The core idea is to establish a time-consistent mapping between "control stick control behavior" and "aircraft attitude response" to reflect the causal coupling characteristics of "operation-response"; the construction of the dynamic coupling features for operation response includes:

[0063] Calculate the integral of the overall manipulation amount within the predetermined forward window, and use this integral as the cumulative manipulation intensity; the specific expression is as follows:

[0064] ;

[0065] The cumulative manipulation intensity up to time t (the cumulative manipulation intent intensity over a period of time). Let hpos_sum(t) = col_hpos_capt(t) + col_hpos_fo(t) be the total manipulation amount for time period j. Here, it should be noted that hpos_sum(t) = col_hpos_capt(t) + col_hpos_fo(t) specifically refers to the "current time point," used to define the value of the total manipulation amount at a specific moment. It describes the characteristics of a single moment. j is the "time point variable within the traceback window", used to iterate and calculate the "comprehensive manipulation amount of the δ moments before time point t (j = t-δ, t-δ+1, ..., t)". It is a traversal representation of all moments within a time window, where δ represents the length of the traceback window (preferably δ=5) and Δt is the sampling interval (preferably Δt=0.125s).

[0066] The integrated control inputs, combined with a first-order hysteresis system model, yield the predicted pitch angle response value; the discrete form of the first-order hysteresis system model is as follows:

[0067] ;

[0068] ;

[0069] in, Let t be the pitch angle predicted response value. Let K be the predicted response value of the pitch angle at time t-1, and K be the gain coefficient (K=0.6, adjusted based on sample experience). The lag factor is preferred (0.75). The operation response lag time constant (preferably τ) r =0.375s, which is 3 sampling periods. The sampling interval for QAR data. The total manipulation amount at time t

[0070] Calculate the actual pitch angle difference. The difference between the predicted pitch angle response value and the predicted pitch angle response value The residuals between; specific residuals The expression is:

[0071] ;

[0072] The smaller the residual, the better the control and response match, and the smoother the leveling operation.

[0073] Multiplying the residual by the cumulative manipulation intensity yields the dynamic coupling characteristics of the operation response. The specific expression used to capture abnormal manipulation actions (such as premature or delayed leveling) is:

[0074] ;

[0075] This achieves the coupling of "cumulative manipulation intensity" and "manipulation-response residual", highlighting the manipulation response characteristics during the flattening phase;

[0076] The formula for calculating the dynamic operation frequency factor is as follows:

[0077] ;

[0078] ;

[0079] in, For dynamic operation frequency factor; for The total manipulation amount at each time step; N is the total number of time steps to be summed, accumulated from 1 to N; This is an indicator function; if the condition inside the square brackets is true, I = 1; otherwise, I = 0. The dynamic threshold (the dynamic threshold reflects the true distribution of "active operations" at different stages by the median offset of the number of operations within the sliding window before and after the current sample point. It is sensitive to flight status, and the threshold is low when there are no large operations within the window, making it easy to capture micro-manipulations; secondly, it is robust to abnormal operations, and uses the median and MAD instead of the mean to filter out the influence of extreme values; finally, it is highly interpretable, and is essentially a local adaptive activity detection); median is the median function; MAD is the median absolute deviation function, which measures the degree of fluctuation. In the time window The overall manipulation quantity; λ is the adjustment factor, generally taken as 1~2, used to control sensitivity; The width of a sliding window is typically 3 to 5.

[0080] The absolute value of hpos_sum within the statistics window exceeds The ratio of the number of samples in the window to the total number of samples in the window is . The larger the value, the more active the manipulation, and the more likely it is to be in a critical stage of leveling off;

[0081] The feature set (including the aforementioned: dynamic coupling features of operation response, cumulative control strength, dynamic operation frequency factor, comprehensive control amount, pitch angle difference features, descent rate difference features, pitch angular velocity, and radio altitude) is input into a pre-trained improved random forest model, and the initial flattening label sequence is output (each timestamp corresponds to a label: 1 indicates "flattening", 0 indicates "non-flattening").

[0082] Model parameter settings: Number of decision trees n_estimators=200 (determined through cross-validation to balance model accuracy and training efficiency), each decision tree is trained independently, and the final result is output through voting (ensemble learning reduces the risk of overfitting).

[0083] Node splitting criterion: Traditional information gain IG(S,A) only considers the discriminative power of features. For example, first, the information entropy I(S) of the node sample set S is calculated, defined as:

[0084] ;

[0085] Where C represents the number of categories (C=2), corresponding to the "unflattened" and "flattened" labels respectively. For the node The proportion of samples of each class;

[0086] Secondly, for candidate feature A, calculate its conventional information gain IG(S, A):

[0087] ;

[0088] Where Values(A) represents all possible values ​​of feature A. This is a subset of samples where feature A takes the value v.

[0089] The improved random forest model uses weighted information gain as the splitting criterion when splitting decision tree nodes. This weighted information gain is calculated by weighting the traditional information gain of the features, historical weight factors, and dynamic operation frequency factors. The formula for the weighted information gain is:

[0090] ;

[0091] in, The weighted information gain of feature A on the sample set S. For the traditional information gain of feature A on sample set S, This is the historical weighting factor for feature A, initially set to 1, and iteratively adjusted during forest construction based on the splitting performance of each feature in previous trees. The dynamic operation frequency factor is used to expand the information gain and reflect the frequency of manipulation actions in the sample.

[0092] Branch growth control: The improved random forest model adopts an adaptive pruning strategy during the branch growth stage of the decision tree;

[0093] To avoid overfitting due to excessively deep decision tree branches, the adaptive pruning strategy is based on local entropy difference: the average information entropy of the child nodes generated when a node splits. Information entropy of parent node reduction amount ,in, Less than the preset threshold When ε=0.01, terminate the current branch expansion. This split contributes very little to distinguishing between "flattened / non-flattened". Continuing to split will introduce noise, so terminate the current branch.

[0094] The initial flattening marker sequence is subjected to neighbor interval fusion, short-term false detection filtering and start time anchoring to obtain the flattening time stamp;

[0095] In this embodiment, the initial labels output by the model may contain "short-term false detections" or "interval breaks," which need to be optimized through rules. The specific process of neighbor interval fusion, short-term false detection filtering, and initial time anchoring is as follows:

[0096] The neighboring interval fusion: For continuous intervals marked as 1 in the initial flattening mark sequence, if the interval between any two intervals does not exceed one QAR data sampling period and at least one interval lasts for more than two QAR data sampling periods, they are merged into one flattening event interval to avoid interval breakage caused by small intervals.

[0097] The short-term false detection filtering: For all flattened event intervals after fusion, if there are multiple intervals and their intervals are more than one QAR data sampling period, and the duration of each interval does not exceed two QAR data sampling periods, then all these intervals will be filtered out.

[0098] Starting time anchoring and output: In the flattened event interval after fusion and filtering, the starting timestamp of the first interval is selected as the final flattening output;

[0099] Comparative Experiment Design and Results Analysis: Three comparative models were selected for experiments: Random Forest (RF), Support Vector Machine (SVM), and Decision Tree (DT). Model performance was comprehensively evaluated using the following metrics: Accuracy, Precision, Recall, and F1 score. The results are shown in Table 1.

[0100] Table 1

[0101] Model Accuracy Precsion Recall F1 score RF 81.18 81.61 79.78 83.53 SVM 90.59 90.12 94.81 85.88 DT 86.47 86.39 86.90 83.88 Algorithm of this invention 96.24 96.23 96.47 96.03

[0102] The results show that the present invention leads in all four indicators. Accuracy is improved by 5.65 percentage points compared with the second best SVM, F1-score is improved by 10.15 percentage points, and the leveling recognition ability is stronger. The difference between Precision and Recall is only 0.24 percentage points, which is much lower than SVM's 4.69 percentage points, and is more in line with the "precise recognition" requirements of flight training.

[0103] It should be noted that the above-described automatic identification method for leveling based on an improved random forest model is merely an example of the division of the functional modules. In practical applications, the functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be further decomposed or combined. For example, the modules in the above embodiments can be merged into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are merely for distinguishing the various modules or steps and are not considered as an improper limitation of the present invention.

[0104] Those skilled in the art will recognize that the modules and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. The programs corresponding to the software modules and method steps can be placed in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art. To clearly illustrate the interchangeability of electronic hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the invention.

[0105] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0106] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0107] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for automatic identification during leveling based on an improved random forest model, characterized in that, The method includes the following steps: Acquire QAR data from the flight training simulator; the QAR data includes radio altitude, pitch angle, rate of pitch change, rate of descent, training tags, and pitch direction control values ​​for the left and right seat stick positions. The QAR data is preprocessed and a feature set is constructed. The preprocessing includes data cleaning and sample construction. The sample construction includes merging the pitch direction control values ​​of the left and right side sticks to generate a comprehensive control value. The features in the feature set include dynamic coupling features of operation response constructed based on the cumulative control intensity and residuals of the comprehensive control quantity, and dynamic operation frequency factor calculated based on the proportion of samples in the comprehensive control quantity that exceed the dynamic threshold within the sliding window; wherein, the residuals are derived based on the difference between the actual pitch angle difference value and the pitch angle predicted response value. The feature set is input into a pre-trained improved random forest model, and the output is an initial flattened label sequence; wherein, the improved random forest model uses weighted information gain as the splitting criterion when splitting decision tree nodes, and the weighted information gain is calculated by weighting the traditional information gain of the feature, the historical weight factor and the dynamic operation frequency factor; the improved random forest model adopts an adaptive pruning strategy during the branch growth stage of the decision tree. The formula for calculating the dynamic operation frequency factor is as follows: ； ； in, For dynamic operation frequency factor, for The total manipulation amount at each time step, where N is the total number of time steps to be summed, accumulated from 1 to N. This is an indicator function; if the condition within the square brackets is true, I = 1; otherwise, I = 0. Here, `median` is the dynamic threshold, `median` is the median function, and `MAD` is the median absolute deviation function. In the time window The overall manipulation amount, where λ is the adjustment factor. Slide window width; The initial flattening marker sequence is subjected to neighbor interval fusion, short-term false detection filtering and start time anchoring to obtain the flattening time stamp.

2. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The construction of the dynamic coupling feature of the operation response includes: Calculate the integral of the overall manipulation amount within the predetermined forward window, and use this integral as the cumulative manipulation intensity; The integrated control inputs, combined with a first-order hysteresis system model, yield the pitch angle prediction response value. Calculate the residual between the actual pitch angle difference and the predicted pitch angle response value; Multiplying the residual by the cumulative manipulation intensity yields the dynamic coupling characteristics of the operation response.

3. The automatic identification method for leveling based on an improved random forest model according to claim 2, characterized in that, The discrete form of the first-order lag system model is as follows: ； ； in, Let be the predicted response value of the pitch angle at time t. Let K be the predicted response value of the pitch angle at time t-1, and K be the gain coefficient. The lag coefficient, The operation response lag time constant, The sampling interval for QAR data. Let t be the total manipulation quantity at time t.

4. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The weighted information gain formula is: ； in, The weighted information gain of feature A on the sample set S. For the traditional information gain of feature A on sample set S, The historical weighting factor for feature A.

5. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The adaptive pruning strategy is based on local entropy difference: when the decrease in the average information entropy of the child nodes generated by node splitting is less than the decrease in the information entropy of the parent node, the current branch expansion is terminated.

6. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The specific process of neighbor interval fusion, short-term false detection filtering, and start-time anchoring is as follows: The neighboring interval fusion: For consecutive intervals marked as 1 in the initial flattening mark sequence, if the interval between any two segments does not exceed one QAR data sampling period and at least one segment has a duration greater than two QAR data sampling periods, they are merged into one flattening event interval; The short-term false detection filtering: For all flattened event intervals after fusion, if there are multiple intervals and their intervals are more than one QAR data sampling period, and the duration of each interval does not exceed two QAR data sampling periods, then all these intervals will be filtered out. Starting time anchoring and output: In the flattened event interval after fusion and filtering, the starting timestamp of the first interval is selected as the final flattening output.

7. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The data cleaning process involves identifying and discarding invalid QAR data with incomplete flight segment information or abnormal parameter jumps, while retaining valid QAR data.

8. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The sample construction includes calculating the first-order difference features of pitch angle and descent rate based on valid data to obtain pitch angle difference features and descent rate difference features, and deleting the first row of null values ​​generated by the difference operation; The generated pitch angle difference feature and descent rate difference feature are used to replace the original pitch angle and descent rate parameters, respectively, as dynamic features; Based on the dynamic characteristics and the comprehensive manipulation amount, the manually labeled starting time of flattening is located as a key sample, and the first N rows of the key sample are taken as a reference sample reflecting the state before flattening. The key samples and the benchmark samples are filtered for duplicate samples through conditional judgment and then added to the feature set.

9. The automatic identification method for leveling based on an improved random forest model according to claim 1, characterized in that, The training labels are manually labeled starting times for leveling. The labeling rules for manual labeling include: the first timestamp where the pitch angle begins to rise continuously and the descent rate begins to decrease continuously is taken as the starting time for leveling and labeled as 1, and the remaining timestamps are labeled as 0.

Images