A real-time maneuver recognition method and apparatus based on early time series classification

Abstract

This invention provides a real-time maneuver recognition method and apparatus based on early time series classification, belonging to the field of intelligent recognition technology. The method first constructs the flight characteristics of an aircraft, then uses a maneuver unit window to divide complete maneuvers into maneuver units, and extracts the temporal features of these maneuver units. Based on the temporal features of the maneuver units, an ensemble learning model, CatBoost, is used to identify the maneuver units. The parameters of the CatBoost ensemble learning model are adjusted to obtain an initial CatBoost ensemble learning model. The contribution of the temporal features used by the initial CatBoost ensemble learning model is evaluated, and the dimensionality is reduced based on the contribution evaluation results to obtain a lightweight model corresponding to the initial CatBoost ensemble learning model. Real-time maneuver recognition is performed based on the lightweight model and the maneuver window. The real-time performance and interpretability of the algorithm of this invention meet practical requirements.

Description

Technical Field

[0001] This application belongs to the field of intelligent recognition technology, and in particular relates to a real-time motor recognition method and device based on early time series classification. Background Technology

[0002] In air combat decision-making, accurately and quickly identifying enemy aircraft maneuvers is crucial for effectively assessing the air combat situation and capturing enemy tactical intentions. In real-world scenarios, identifying maneuvers only after they have been fully executed loses its practical significance; therefore, it is necessary to identify the complete maneuver type by observing only a portion of the enemy aircraft's flight data. The identification of fighter jet maneuvers is a key research area in air combat both domestically and internationally, and also a challenge in the field of artificial intelligence.

[0003] Currently, scholars in the field of air combat both domestically and internationally have proposed various methods for fighter jet maneuver recognition: Expert experience, which uses expert knowledge to build an expert knowledge base and matches maneuvers according to rules to identify them; however, this method is limited by expert knowledge, the rule base may be incomplete, and it cannot identify maneuvers in real time; Multidimensional dynamic time warping method, which calculates the distance between the maneuver sequence to be identified and the maneuver database, and selects the closest maneuver category as the category of the maneuver sequence to be identified; however, this method has low computational efficiency, cannot identify in real time, and the recognition accuracy depends on the setting of the maneuver database; Deep learning method, which uses the LSTM algorithm to capture the relationship between data in the time series to identify fighter jet maneuvers; however, deep learning algorithms have poor interpretability, which is not conducive to human-machine mutual trust in practical applications. Summary of the Invention

[0004] To address the technical problems of insufficient real-time performance and interpretability in related algorithms, this invention provides a real-time maneuver recognition method and apparatus based on early time series classification. The technical solution is as follows:

[0005] Firstly, a real-time maneuver recognition method based on early time series classification is provided, including:

[0006] Based on the flight characteristics of the constructed aircraft, the complete maneuver is divided into maneuver units using a maneuver unit window, and the temporal features of the maneuver units are extracted.

[0007] Based on the temporal features of the maneuvering unit, the ensemble learning model CatBoost is used to identify the maneuvering unit. The parameters of the ensemble learning model CatBoost are adjusted to obtain the initial ensemble learning model CatBoost.

[0008] The contribution of the temporal features used in the initial ensemble learning model CatBoost is evaluated, and then the dimensionality is reduced based on the contribution evaluation results to obtain a lightweight model corresponding to the initial ensemble learning model CatBoost.

[0009] Real-time maneuver recognition is achieved using a lightweight model and a maneuver window.

[0010] Among them, real-time maneuver recognition based on a lightweight model and a maneuver window includes:

[0011] Complete motion identification is performed by combining the motion window and motion unit identification results: when multiple consecutive motion unit identification results are consistent, the complete motion type is determined to be the type of motion unit.

[0012] Furthermore, before dividing a complete maneuver into maneuver units using a maneuver unit window based on the flight characteristics of the constructed aircraft, the method further includes:

[0013] Obtain the motor recognition dataset;

[0014] The acquired motor recognition dataset is divided into a training set, a validation set, and a test set;

[0015] The training set, validation set, and test set are divided and transformed using coordinate system transformation and noise filtering to obtain standard data;

[0016] The flight characteristics of the aircraft are constructed based on standard data.

[0017] Among them, flight characteristics include the rate of change of flight speed, pitch angle, heading angle and attitude angle, where attitude angle is either pitch angle or heading angle.

[0018] The time series characteristics include mean, variance, maximum value, minimum value, range, standard deviation, and coefficient of variation.

[0019] The process of acquiring the mobile unit window is as follows:

[0020] The validation set data is processed using a window size of Ts and Tw, dividing the longer time series into multiple subsequences. Each subsequence is a mobile unit window, with Tw ranging from 20 to 54 and Ts being half of Tw.

[0021] Secondly, a real-time motion recognition device based on early time series classification is provided, comprising:

[0022] The segmentation unit is used to divide the complete maneuver into maneuver units based on the flight characteristics of the constructed aircraft, and extract the temporal features of the maneuver units.

[0023] The training unit is used to identify the maneuvering unit using the CatBoost ensemble learning model based on the temporal features of the maneuvering unit. The parameters of the CatBoost ensemble learning model are adjusted to obtain the initial CatBoost ensemble learning model.

[0024] The evaluation unit is used to evaluate the contribution of the temporal features used in the initial ensemble learning model CatBoost, and then the dimensionality is reduced based on the contribution evaluation results to obtain the lightweight model corresponding to the initial ensemble learning model CatBoost.

[0025] The recognition unit is used for real-time maneuver recognition based on a lightweight model and a maneuver window.

[0026] The identification unit is specifically used for:

[0027] Complete motion identification is performed by combining the motion window and motion unit identification results: when multiple consecutive motion unit identification results are consistent, the complete motion type is determined to be the type of motion unit.

[0028] To address the real-time performance issue of the algorithm: This invention designs a two-layer recognition mode. First, the complete maneuver is divided into multiple shorter maneuver units through a time window, and the temporal features of the maneuver units are extracted. Then, an ensemble learning model is used to identify the maneuver units. Real-time performance is ensured by monitoring the maneuver units. Finally, the early time series classification decision points are determined through the maneuver window to identify the complete maneuver. To address the interpretability issue of the algorithm: This invention uses an interpretable model to explain the model through the contribution of features. Then, features are selected according to their contribution for ablation experiments to obtain the optimal feature combination.

[0029] During maneuver recognition, this invention can complete the recognition process within a portion of the time following the start of the maneuver. Simultaneously, it interprets the contribution of the features used by the model, improving the model's interpretability and ensuring user trust in the model's output. Experiments have demonstrated that this invention achieves performance improvements in metrics such as accuracy, F1 score, recall, precision, and recognition latency. Attached Figure Description

[0030] Figure 1 A flowchart of a real-time maneuver recognition method based on early time series classification provided by the present invention;

[0031] Figure 2 The experimental results for finding the optimal time window provided by this invention. Detailed Implementation

[0032] The present application will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0033] Real-time maneuver identification aims to identify enemy aircraft maneuvers as early as possible, constituting an early time series classification task. Early time series classification refers to classifying the entire sequence based on data from the early stages of observation. It has two key objectives: high-precision classification and early classification. However, these two objectives are inherently contradictory: earlier classification means less data and lower accuracy. Therefore, early time series classification requires ensuring real-time classification without sacrificing accuracy. This invention, based on the concept of early time series classification, employs a two-layer algorithm structure for maneuver identification. It uses an ensemble learning model for classification, introduces an interpretable model for interpretation, and finally determines the minimum prediction length through a maneuver window for real-time identification.

[0034] To address the real-time performance issue of the algorithm: This invention designs a two-layer recognition mode. First, the complete maneuver is divided into multiple shorter maneuver units through a time window, and the temporal features of the maneuver units are extracted. Then, an ensemble learning model is used to identify the maneuver units. Real-time performance is ensured by monitoring the maneuver units. Finally, the early time series classification decision points are determined through the maneuver window to identify the complete maneuver. To address the interpretability issue of the algorithm: This invention uses an interpretable model to explain the model through the contribution of features. Then, features are selected according to their contribution for ablation experiments to obtain the optimal feature combination.

[0035] This invention provides a real-time aircraft maneuver recognition method based on early time series classification. While there is extensive research on aircraft maneuver recognition methods, most algorithms are applied to offline scenarios and cannot perform maneuver recognition in real time; furthermore, existing models lack interpretability, and in the field of air combat decision-making, opaque decisions expose decision-makers to risk. This invention offers an innovative maneuver recognition method that utilizes an early time series classification algorithm to perform maneuver recognition. It achieves real-time recognition by dividing the complete maneuver into smaller segments through time windows and makes the model interpretable through feature contribution analysis.

[0036] In this invention, a complete maneuver is divided into maneuver units of equal size using a time window, and the temporal relationships within these units are extracted. The CatBoost ensemble learning model is then used to identify these maneuver units. Dividing a complete maneuver into maneuver units makes maneuver recognition more real-time. Secondly, the contribution of features used by the interpretable algorithm library Shap is evaluated using the feature contribution to obtain a lightweight model, making maneuver recognition interpretable. Finally, the maneuver unit identification results are windowed; only when multiple consecutive maneuver unit identification results are consistent is the complete maneuver type considered a maneuver unit type, ensuring the accuracy of the identification results.

[0037] This invention provides a real-time maneuver recognition method based on early time series classification, such as... Figure 1 As shown, it includes the following steps:

[0038] Step 1: Obtain the maneuver recognition dataset, simulate various maneuvers using the DCS World simulator, and output the 3D coordinate data A of the maneuvers using the data logging software Tacview 1.9.0. LLA = [Longitude, latitude, altitude], the sampling frequency in the experiment was 20 Hz, that is, the interval between each sampling point was 0.05 seconds. The maneuvers identified by this invention are shown in Table 1, including the following nine types: somersault, climb, dive, turn, S-bend, spiral, taildive, sharp turn, and sledgehammer maneuver. Five maneuvers are selected from the above nine types and combined to form different maneuver combinations.

[0039] Step 2: Divide the acquired motor recognition dataset into training set Dtr, validation set Dva, and test set Dte in a 3:1:1 ratio, as shown in Table 1.

[0040] This step facilitates subsequent training of the CatBoost ensemble learning model using the training set data, selection of the optimal time window and optimal features on the validation set data, and testing of the CatBoost ensemble learning model's performance on the test set data.

[0041] Table 1. Motor recognition dataset

[0042]

[0043] Step 3: Transform the divided training, validation, and test sets using coordinate system transformation and noise filtering to obtain standard data A. ENU = [X, Y, Z].

[0044] In air combat, if the radar only acquires the target's longitude (X), latitude (Y), and altitude (Z) coordinates, and the acquired data may contain noise, this will be very detrimental to real-time identification of the enemy aircraft's maneuvers. Therefore, it is necessary to standardize the maneuver identification data. First, coordinate system transformation is used to transfer all maneuver identification data to a unified coordinate system, and then Kalman filtering is used to reduce noise in the maneuver identification data.

[0045] Step 4: Based on domain knowledge, construct the aircraft's flight characteristics according to standard data. The flight characteristics include the rate of change of flight speed, pitch angle, heading angle, and attitude angle, where the attitude angle is either pitch angle or heading angle.

[0046] 1. Based on standard data A ENU = [X, Y, Z], calculate the velocity [v] of the three-dimensional coordinate axes using the following formula. x ,vy ,v z ]:

[0047]

[0048] Where t is the sampling point, Δt is the sampling time, and X = [x1, x2, ... x t ,...x n ], Y = [y1, y2, ... y t ,...y n Z = [z1, z2, ... z] t ,...z n ], v x =[v x1 ,v x2 ,...v xt ,...v xn ], v y =[v y1 ,v y2 ,...v yt ,...v yn ], v z =[v z1 ,v z2 ,...v zt ,...v zn ].

[0049] 2. Calculate the aircraft's flight speed based on the velocities along the three-dimensional coordinate axes:

[0050]

[0051] Among them, v x ,v y ,v z Three-axis velocity;

[0052] 3. Calculate the aircraft's pitch angle based on the three-axis speeds:

[0053]

[0054] Among them, v x ,v y ,v z Three-axis velocity;

[0055] 4. Calculate the aircraft's heading angle based on the three-axis velocities:

[0056]

[0057] Among them, v x ,v y For the x-axis and y-axis velocities;

[0058] 5. Calculate the rate of change of attitude angle based on the three-axis velocity:

[0059]

[0060] Where angle is the pitch angle or heading angle.

[0061] Step 5: Divide the complete maneuver into time windows to obtain maneuver units, and extract the temporal features of the maneuver units.

[0062] The above feature construction can reflect the aircraft's flight status at the sampling point, but it cannot reflect the temporal characteristics of the aircraft's maneuvers. For time series classification problems, defining a time window to extract features can effectively capture the characteristics of data over a period of time. The extracted temporal features include: mean, variance, maximum value, minimum value, range, standard deviation, and coefficient of variation—seven-dimensional temporal features.

[0063] Since the size of the time window is a hyperparameter, an optimal time window needs to be found using the results on the validation set Dva. Using the ensemble learning model CatBoost, the optimal time window is selected by comparing the accuracy of identifying maneuver units under different window conditions for five different maneuver combinations.

[0064] In this invention, a window size of Ts and Tw is used to process the validation set Dva data, dividing the longer time series T into multiple subsequences:

[0065] X = {X i ,i=1,2,3,...,m},

[0066] in, During the experiment, based on experience, the sampling time window (Tw) was designed to range from 20 to 54 points, with Ts being half the size of Tw. Different Tw values ​​were used to capture different states of the time series, and the optimal time window was ultimately selected. For example, the optimal time window for dividing maneuvering units was found experimentally. Using the CatBoost model, the accuracy of identifying maneuvering units under different window conditions for five combinations of maneuvers was compared to select the optimal time window. Figure 2 As shown, the design time window size ranges from 20 to 54, with two windows added per round for analysis results. To meet the real-time and accuracy requirements of maneuver recognition, a maneuver segment with a time window size of 40 is selected as the maneuver unit, and a new time window is divided each time 20 sampling points are entered. Calculations show that the time length of each maneuver unit is 2 seconds, and the time window is re-divided every 1 second.

[0067] The data from the aforementioned time window constitutes a maneuvering unit. Through steps 2 to 5, the originally independent sampling points are divided into maneuvering units, and the target training set Dtr, target validation set Dva, and target test set Dte are obtained.

[0068] Step 6: Based on the temporal features of the maneuvering units, the ensemble learning model CatBoost is used to identify the maneuvering units. The parameters of the ensemble learning model CatBoost are adjusted to obtain the initial ensemble learning model CatBoost.

[0069] CatBoost is a mainstream ensemble learning model, known for its high prediction speed and accuracy. CatBoost uses the GBDT framework based on symmetric decision trees.

[0070] First, the parameters in the CatBoost ensemble learning model need to be adjusted to change it from a weak classifier to a strong learner. The parameter adjustment process for the CatBoost model is as follows:

[0071] Let's define a motor recognition dataset. x i It refers to any one of the following maneuvers: somersault, climb, dive, turn, S-bend, spiral, taildive, sharp turn, and sledgehammer maneuver. i Indicates the motor recognition result, when y i When y = 0, it indicates that the maneuver does not belong to this category; when y = 0, it indicates that the maneuver does not belong to this category. i =1 indicates that the maneuver belongs to this category.

[0072] The first step is to initialize the weak classifier as follows:

[0073]

[0074] F0(x) is the initial weak classifier, which will become a powerful classifier after multiple iterations.

[0075] The second step, for the m-th iteration m = 1, 2, ..., M:

[0076] (1) For each sample i, i = 1, 2, ..., n, use the loss function L(y i F m (x)) Calculate the negative gradient value and approximate it with the residual:

[0077]

[0078] (2) Use the residual obtained in the previous step Training the next tree f m (x), let J be the number of leaf nodes in tree m, and R be the leaf region. mj j = 1, 2, ..., J

[0079] (3) For j = 1, 2, ..., J, calculate the best-fit value for the leaf region:

[0080]

[0081] Since solving the above equation is difficult, its approximate value is calculated:

[0082]

[0083] The learner is updated using the above formula, as follows:

[0084]

[0085] The third step is to finally obtain the strong learner, which is the initial CatBoost model:

[0086]

[0087] The above process is applied to binary classifiers, but in maneuver recognition, there are K classes. To address this, ONE-HOT encoding is used to first convert each class into a K-dimensional vector. K trees are trained simultaneously in each round. Then, the loss function is modified to generate probabilities, determining the probability that a maneuver belongs to a certain class. Ultimately, this achieves the classification of multiple maneuvers and completes the recognition of maneuver units.

[0088] Step 7: Use the interpretable algorithm library Shap to interpret the initial ensemble learning model CatBoost, obtain the contribution of each temporal feature to the output of the initial ensemble learning model CatBoost, and then perform dimensionality reduction based on the contribution to obtain the lightweight model corresponding to the initial ensemble learning model CatBoost. This lightweight model is the optimized ensemble learning model CatBoost.

[0089] Shap is an interpretable algorithm library that calculates the Shapley value of each temporal feature to determine its contribution to the output of the initial ensemble learning model, CatBoost. The formula for calculating the contribution is:

[0090]

[0091] Where g is the interpretable algorithm library Shap, M is the dimension of the input features, and t is the presence or absence of the corresponding time-series feature. For a constant term, This represents the Shapley value for each flight feature. Since the ensemble learning model Catboost used is a tree model, all temporal features are present, and the formula can be rewritten as:

[0092]

[0093] The difficulty of the above formula lies in finding... In the tree model, for a given feature i, the Shapley value needs to be calculated for all possible feature combinations, and then a weighted sum is performed. Through this calculation, the contribution of each temporal feature to maneuver recognition can be determined. Then, ablation experiments are conducted on the initial ensemble learning model CatBoost to select several important temporal features for maneuver recognition, ensuring the interpretability of the initial ensemble learning model CatBoost. Table 2 shows the optimal temporal feature combinations under different dimensions, revealing that the most important temporal feature of a maneuver is the change in the aircraft's attitude angles during the maneuver.

[0094] Table 2 Optimal Combinations of Time-Series Features in Different Dimensions

[0095]

[0096] Step 8: Perform real-time maneuver recognition based on a lightweight model and a maneuver window.

[0097] Maneuver windows are used to determine early time series classification decision points, and the complete maneuver type is output in real time. Assuming the maneuver sequence is divided into n maneuver units, the maneuver sequence can be represented as M = {Mu1, Mu2, Mu3, ..., Mu...}. n}, where M is the identification result after the maneuver sequence is split into maneuver units, and Mu is the identified maneuver unit. The maneuver window can be represented as Mw = [Mu k Mu k+1 , ..., Mu k+m ], where Mw is the maneuver window, k is the kth identified maneuver unit, and m is the length of the maneuver window. A maneuver is considered to have occurred or changed only when all the maneuvers in Mw are of the same type.

[0098] For early time series classification, the maneuver window is the minimum prediction length, i.e., the decision point for early time series classification. Designing a maneuver window can eliminate the impact of misidentified maneuver units on the overall maneuver recognition, and ensure that the recognition result does not arbitrarily change during a maneuver execution, thus improving the accuracy and real-time performance of maneuver recognition. In real-time maneuver recognition, the size of the maneuver window is a hyperparameter that can be determined experimentally, using the accuracy of the maneuver sequence on the validation set Dva to determine the size of the maneuver window.

[0099] For example, the identification results of the maneuver units in Table 1 are shown in Table 3 below.

[0100] Table 3. Results of Mobile Unit Identification

[0101] Mobile combination accuracy accuracy Recall rate F1 Mobile Combination 1 0.9373 0.9152 0.9141 0.9131 Mobile Combination 2 0.9154 0.9094 0.9266 0.9167 Mobile Combination 3 0.9294 0.9125 0.9110 0.9105 Mobile Combination 4 0.9390 0.9374 0.9433 0.9397 Mobile Combination 5 0.9162 0.9133 0.9007 0.9051

[0102] As shown in Table 3, the present invention has high accuracy in identifying maneuver units in various maneuver combinations, and can effectively capture the characteristics of maneuver execution and accurately identify maneuver actions.

[0103] A maneuver window is added for real-time maneuver recognition. Experiments on the target validation set Dva showed that a maneuver window size of 6 was selected, meaning the invention can recognize the maneuver at 7 seconds after its start. Table 4 shows the latency evaluation metrics for this invention.

[0104] Table 4. Complete Maneuver Recognition Latency Rate

[0105] Motor Name Complete time series execution time Latency somersault motor 40S 0.1750 Climb maneuver 25S 0.2800 dive maneuver 25S 0.2800 circling maneuver 60S 0.1167 Break S-Mobility 25S 0.2800 Spiral Maneuver 60S 0.1167 Tail-run maneuver 45S 0.1556 sharp turn maneuver 40S 0.1750 Hammer Mobility 35S 0.2000

[0106] Step 9: During testing, the data from the target test set Dte is fed into the real-time maneuver recognition model, which consists of the optimized CatBoost model and the maneuver window. During maneuver recognition, accuracy (acc) and the overall evaluation score F1 (F1_score) are used as the primary evaluation metrics. To more comprehensively evaluate the model's performance, recall (R) and precision (Pre) are used as auxiliary evaluation metrics. Furthermore, since this invention is based on an early time series classification method, the earlyness of recognition is also an important evaluation metric; therefore, the earlyness of recognition is evaluated using the recognition delay rate (D).

[0107] Accuracy, F1 score, recall, precision, and recognition latency are represented as follows:

[0108] Accuracy: The percentage of all maneuver data that is correctly identified.

[0109] F1_score: The harmonic mean of precision and recall. The closer the F1_score is to 1, the more accurate the classification.

[0110] Recall: The proportion of data correctly identified for a particular maneuver type out of all data for that maneuver type.

[0111] Precision: The proportion of data correctly identified as belonging to this maneuver type out of all data identified as belonging to this maneuver type;

[0112] Delay Rate: The proportion of the time taken to identify the maneuver within the total maneuver time. The smaller the proportion, the earlier the maneuver can be identified at the beginning of the maneuver.

[0113] The present invention is compared with other time series models. Taking combination one as an example, the comparison results of the present invention with other models are shown in Table 5 below.

[0114] Table 5 Comparison of Complete Mobility Recognition Results

[0115]

[0116]

[0117] It is evident that this model achieves higher classification accuracy, precision, recall, and F1 score than other methods, and also exhibits low latency, short runtime, and strong real-time performance.

[0118] Based on the above experiments, this invention utilizes readily available three-dimensional coordinate information of enemy aircraft in air combat to design an interpretable real-time maneuver recognition algorithm based on early time series classification, enabling real-time identification of enemy aircraft maneuvers. The algorithm is interpretable, employing feature contribution analysis to ensure user trust in the algorithm's output. The algorithm achieves 90% accuracy across nine different combinations of maneuvers, with a recognition time of only 7 seconds, averaging only 20% of the total maneuver execution time. This invention demonstrates high accuracy, strong real-time performance, and good interpretability, making it highly practical.

[0119] This invention also provides a real-time maneuver recognition device based on early time series classification, comprising:

[0120] The segmentation unit is used to divide the complete maneuver into maneuver units based on the flight characteristics of the constructed aircraft, and extract the temporal features of the maneuver units.

[0121] The training unit is used to identify the maneuvering unit using the CatBoost ensemble learning model based on the temporal features of the maneuvering unit. The parameters of the CatBoost ensemble learning model are adjusted to obtain the initial CatBoost ensemble learning model.

[0122] The evaluation unit is used to evaluate the contribution of the temporal features used in the initial ensemble learning model CatBoost, and then the dimensionality is reduced based on the contribution evaluation results to obtain the lightweight model corresponding to the initial ensemble learning model CatBoost.

[0123] The recognition unit is used for real-time maneuver recognition based on a lightweight model and a maneuver window.

[0124] The identification unit is specifically used for:

[0125] Complete motion identification is performed by combining the motion window and motion unit identification results: when multiple consecutive motion unit identification results are consistent, the complete motion type is determined to be the type of motion unit.

[0126] This invention first performs coordinate system normalization and noise filtering on the maneuver recognition dataset to obtain standardized maneuver data. Second, using domain knowledge and a sliding window, it processes the limited three-dimensional coordinate information of the maneuver time-series data to construct a large number of maneuver time-series features, and divides complete maneuvers into maneuver units. Subsequently, an ensemble learning model selects the optimal time window to divide the maneuver units and performs recognition. Then, the Shap interpretable algorithm library is used to interpret the feature contribution of the ensemble learning model, obtaining the optimal feature combination. Finally, real-time maneuver recognition is achieved through the maneuver window.

[0127] This invention demonstrates high accuracy and interpretability in maneuver recognition, solving the problem of real-time aircraft maneuver recognition and providing information support for battlefield situational awareness and decision support.

[0128] The above description merely illustrates the embodiments of this application, and while the description is relatively specific and detailed, it should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Furthermore, any parts not detailed in this application are conventional techniques.

Claims

1. A real-time maneuver recognition method based on early time series classification, characterized in that, include: Based on the flight characteristics of the constructed aircraft, the complete maneuver is divided into maneuver units using a maneuver unit window, and the temporal features of the maneuver units are extracted. Based on the temporal features of the maneuvering unit, the ensemble learning model CatBoost is used to identify the maneuvering unit. The parameters of the ensemble learning model CatBoost are adjusted to obtain the initial ensemble learning model CatBoost. The contribution of the temporal features used in the initial ensemble learning model CatBoost is evaluated, and then the dimensionality is reduced based on the contribution evaluation results to obtain a lightweight model corresponding to the initial ensemble learning model CatBoost. Real-time maneuver identification based on lightweight models and maneuver windows: Complete maneuver identification is performed by combining maneuver windows and maneuver unit identification results: When multiple consecutive maneuver unit identification results are consistent, the complete maneuver type is determined as the type of the maneuver unit, specifically: The complete maneuver is divided into maneuver units of equal size through a time window, and the temporal relationships in the maneuver units are extracted. Then, the ensemble learning model CatBoost is used to identify the maneuver units. Using the interpretable algorithm library Shap, the contribution of features used by the ensemble learning model CatBoost is evaluated, and then dimensionality reduction is performed through feature contribution to obtain a lightweight model, making maneuver recognition interpretable. The maneuver unit recognition results are then windowed, and only when multiple consecutive maneuver unit recognition results are consistent is the complete maneuver type considered to be a maneuver unit type.

2. The method according to claim 1, characterized in that, Based on the flight characteristics of the constructed aircraft, before dividing the complete maneuver into maneuver units using a maneuver unit window, the method further includes: Obtain the motor recognition dataset; The acquired motor recognition dataset is divided into a training set, a validation set, and a test set; The training set, validation set, and test set are divided and transformed using coordinate system transformation and noise filtering to obtain standard data; The flight characteristics of the aircraft are constructed based on standard data.

3. The method according to claim 1, characterized in that, Flight characteristics include the rate of change of flight speed, pitch angle, heading angle, and attitude angle, where the attitude angle is either pitch angle or heading angle.

4. The method according to claim 1, characterized in that, Time series characteristics include mean, variance, maximum value, minimum value, range, standard deviation, and coefficient of variation.

5. The method according to claim 1, characterized in that, The process of acquiring the mobile unit window is as follows: The validation set data is processed using a window size of Ts and Tw, dividing the longer time series into multiple subsequences. Each subsequence is a mobile unit window, with Tw ranging from 20 to 54 and Ts being half of Tw.

6. A real-time motion recognition device based on early time series classification, characterized in that, The apparatus for performing the method of claim 1 includes: The segmentation unit is used to divide the complete maneuver into maneuver units based on the flight characteristics of the constructed aircraft, and extract the temporal features of the maneuver units. The training unit is used to identify the maneuvering unit using the CatBoost ensemble learning model based on the temporal features of the maneuvering unit. The parameters of the CatBoost ensemble learning model are adjusted to obtain the initial CatBoost ensemble learning model. The evaluation unit is used to evaluate the contribution of the temporal features used in the initial ensemble learning model CatBoost, and then the dimensionality is reduced based on the contribution evaluation results to obtain the lightweight model corresponding to the initial ensemble learning model CatBoost. The recognition unit is used for real-time maneuver recognition based on a lightweight model and a maneuver window; The identification unit is specifically used for: Complete maneuver identification is performed by combining the maneuver window and maneuver unit identification results: when multiple consecutive maneuver unit identification results are consistent, the complete maneuver type is determined as the type of maneuver unit, specifically as follows: The complete maneuver is divided into maneuver units of equal size through a time window, and the temporal relationships within the maneuver units are extracted. The ensemble learning model CatBoost is then used to identify the maneuver units. The contribution of the features used by the CatBoost ensemble learning model is evaluated using the interpretable algorithm library Shap, and then dimensionality reduction is performed based on the feature contribution to obtain a lightweight model, making the maneuver identification interpretable. The maneuver unit identification results are then windowed, and only when the identification results of multiple consecutive maneuver units are consistent is the complete maneuver type considered to be a maneuver unit type.

Images