A radar intelligent classification and identification method and system for low-altitude supervision

Abstract

The present application belongs to the field of artificial intelligence and radar signal processing, and relates to a radar intelligent classification and identification method and system for low-altitude supervision, comprising: collecting track data of an aerial flight target by a radar to obtain a track data set; performing feature processing on the track data set to obtain track features; the track features include track curvature, climb rate and hovering index; performing enhancement compression processing on the track features to obtain a mixed feature set, and constructing a training sample; constructing a CNN-LSTM hybrid neural network model with a self-defined effective step length extraction layer; the self-defined effective step length extraction layer is used to eliminate the model performance degradation problem caused by unequal length sequence padding noise; training and verifying the hybrid neural network model to obtain a classification model; processing real-time mixed features by the classification model to obtain target track classification results; and improving the classification accuracy and robustness of radar target identification based on the track level.

Description

TECHNICAL FIELD

[0001] The application relates to the fields of artificial intelligence and radar signal processing, and specifically discloses a radar intelligent classification and identification method and system for low-altitude supervision. BACKGROUND

[0002] Radar is widely used in low-altitude supervision and security fields due to its all-weather, all-day, and high-precision characteristics. In the field of target identification and monitoring, various radar signal processing and classification technologies have been developed, mainly including one-dimensional range imaging, SAR / ISAR imaging, micro-Doppler, radar polarization, and track motion parameter analysis. Among them, the one-dimensional range imaging method is particularly sensitive to target attitude; the SAR / ISAR image processing process is extremely complex, and has the problems of large calculation amount and poor real-time performance; the method based on polarization features needs to rely on a full-polarization radar system, and has the problems of high hardware cost and complex data processing process; the method based on micro-Doppler has the problem of difficult target feature extraction due to weak echo characteristics; and the method combining track motion parameters and deep learning is one of the most effective methods in radar target processing.

[0003] Modern phased array radars can simultaneously track hundreds of targets (such as civil aviation passenger planes, unmanned aerial vehicles, birds, etc.) through electronic scanning technology and form continuous tracks. Combined with track information and artificial intelligence algorithms, the motion trajectory features of the target (such as RCS (radar cross section), speed, acceleration, heading, and trajectory shape, etc.) can be extracted, and target classification and identification can be realized. Compared with traditional identification methods based on images, signal features, or single physical properties, the method of extracting motion trajectory features by combining track information and artificial intelligence algorithms has the advantages of low radar hardware requirements, insensitivity to target shape changes, and low processing complexity; even in scenes with low radar resolution and insufficient information, it can still achieve high identification accuracy.

[0004] Traditional track classification methods mostly rely on a small number of manually selected features (such as average speed, maximum height, etc.), and fail to fully exploit the rich information (such as target maneuverability, motion pattern stability, etc.) contained in the track data, resulting in insufficient feature representation capability and low classification accuracy in complex scenarios. Although deep learning models such as CNN and LSTM can automatically extract features, there is an inherent problem of inconsistent sequence length in the actual processing process. Although the mainstream deep learning frameworks (such as TensorFlow and PyTorch) can solve this problem through padding and masking mechanisms, in some industrial application environments (such as MATLAB platform), there is a lack of native masking function support. When training radar track sequences with variable length, if the padding method is used, redundant noise will be introduced in the training process due to the padding strategy; and in the real-time deployment stage, it is required to process each frame of data without padding, which causes inconsistency between the training and deployment environments. Training on padded sequences will cause the model to learn meaningless padding noise, which seriously affects the performance and generalization ability of the model.

[0005] Therefore, the present application provides a radar intelligent classification and identification method and system for low-altitude supervision, which is suitable for track target identification of radar, ADS-B and other sensors, and aims to improve the radar target identification accuracy and robustness based on track level. SUMMARY

[0006] The present application aims to provide a radar intelligent classification and identification method and system for low-altitude supervision, which solves the problem of improving the classification accuracy and robustness of radar target identification based on track level. The specific scheme is as follows:

[0007] The application discloses a radar intelligent classification and identification method for low-altitude supervision, and relates to the technical field of radar target classification and identification.

[0008] Further, the hovering index is an index quantifying the ability of the target to maintain position stability in three-dimensional space, and the calculation formula of the hovering index is:

[0009] ;

[0010] ;

[0011] ;

[0012] wherein, represents the hovering index; i represents the current time step; represents the proportion of track points with speed lower than the threshold value in the window; η represents a scaling coefficient; represents the sum of variances of three-dimensional coordinates of the target in the window; N is the length of the sliding window; k ∈ [i-N+1, i] represents the index of the time step in the sliding window; and respectively represent the horizontal axis coordinates of the target track at the kth and k-1th time steps; and respectively represent the vertical axis coordinates of the target track at the kth and k-1th time steps; and respectively represent the vertical axis coordinates of the target track at the kth and k-1th time steps; the resultant velocity of the kth time step trajectory; the velocity threshold value for determining the hovering of the trajectory; the indicator function.

[0013] Further, the climbing rate is used to measure the rate of change of the height of the target in the vertical direction, and the calculation formula of the climbing rate is:

[0014] ;

[0015] ;

[0016] wherein, the climbing rate; i represents the current time step; the resultant velocity of the i th time step trajectory; sin represents the sine function; the climbing angle of the i th time step trajectory; arctan represents the inverse tangent function; and respectively represent the height of the i th and i-1 th time step trajectory; and respectively represent the horizontal distance of the i th and i-1 th time step trajectory.

[0017] Further, the trajectory curvature is used to quantify the bending degree of the target trajectory, and the calculation formula of the trajectory curvature is:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] wherein, the trajectory curvature; i represents the current time step; , and respectively represent the horizontal axis, vertical axis and vertical axis coordinates of the i th time step trajectory; , and respectively represent the horizontal axis, vertical axis and vertical axis coordinates of the i-1 th time step trajectory; , and respectively represent the horizontal axis, vertical axis and vertical axis coordinates of the i-2 th time step trajectory; represents the forward vector; represents the backward vector; represents the span vector; represents the vector and The length of the modulus of the cross product.

[0023] Further, the track features are enhanced and compressed to obtain a mixed feature set, including: performing feature enhancement processing on the track features to obtain an enhanced feature dataset; the feature enhancement processing refers to expanding the feature dataset; performing hierarchical sampling on the enhanced feature dataset to obtain a divided dataset; the proportion of targets of each category in a subset of the divided dataset is consistent with the track dataset; performing feature scaling on the features in the divided dataset using Z-Score standardization to obtain a scaled dataset; the data in the scaled dataset is scaled to a normal distribution with a mean of 0 and a variance of 1; performing feature importance evaluation on the scaled dataset, and performing feature compression according to the evaluation results to obtain a compressed dataset, and taking the compressed dataset as the mixed feature set.

[0024] Further, the scaled dataset is evaluated for feature importance, and the compressed dataset is obtained according to the evaluation results, including: destroying the relevance of features and target variables through random forest, evaluating the performance of the mixed neural network model, and obtaining the degree of performance decline; through Lasso regression sparsity solution, some features are sparsely compressed to zero to obtain important features and unimportant features, and based on the values obtained after the important features and unimportant features are compressed, the compression degree is determined; the important features refer to the features that are not compressed to zero, and the unimportant features refer to the features that are compressed to zero; the features of the validation set are replaced and disturbed on the classification model through the importance replacement, the features are evaluated through the volatility of the output performance of the classification model, and the volatility degree is obtained; the difference degree is obtained by quantifying the explanatory power of the features to the category difference through variance analysis; the comprehensive importance index is generated by weighting and comprehensively scoring the performance decline degree, the compression degree, the volatility degree and the difference degree; and the set composed of multiple features whose comprehensive importance index is greater than the preset dynamic threshold is taken as the compressed dataset.

[0025] Further, the hybrid neural network model comprises a sequence input layer, a 1D convolution layer, a batch normalization layer, a first activation function layer, an LSTM layer, a self-defined effective step extraction layer, a first full connection layer, a second activation function layer, a second full connection layer, a Softmax layer and a classification output layer; the sequence input layer is used for inputting the hybrid features; the 1D convolution layer is used for performing convolution processing on the input hybrid features; the batch normalization layer is used for performing normalization processing on the hybrid features after the convolution processing; the first activation function layer is used for performing activation processing on the hybrid features after the normalization processing; the LSTM layer is used for performing long-short time feature extraction on the hybrid features after the activation processing; the self-defined effective step extraction layer is used for performing effective feature extraction on the hybrid features after the long-short time feature extraction and the hybrid features input by the sequence input layer, to obtain effective features; the first full connection layer is used for performing first full connection on the effective features; the second activation function layer is used for performing activation on the effective features after the first full connection; the second full connection layer is used for performing second full connection on the effective features after the activation; the Softmax layer is used for converting the effective features after the second full connection, to obtain a probability distribution of target classification; and the classification output layer is used for outputting a target track classification result based on the probability distribution of target classification.

[0026] Further, the hybrid neural network model is trained and verified by using training samples, to obtain a classification model, comprising: training on a training set, monitoring performance on a verification set, and optimizing and early stopping through hyperparameters, until the performance of the hybrid neural network model on the verification set reaches a cycle termination condition, to obtain a final hybrid neural network model; finally evaluating the final hybrid neural network model by using a test set, to obtain a generalization performance index; the test set refers to training samples that have not participated in any training process; and the final hybrid neural network model with the generalization performance index reaching a generalization performance requirement is taken as the classification model.

[0027] Further, real-time hybrid features are obtained by performing data processing on real-time track data collected by a radar, comprising: performing feature processing on the real-time track data, to obtain real-time track features; performing enhancement and compression processing on the real-time track features, to obtain the real-time hybrid features; further comprising: obtaining a real track type of a target track; and continuously optimizing the classification model based on the real track type and the target track classification result.

[0028] The application further provides a low-altitude monitoring-oriented radar intelligent classification and identification system using the low-altitude monitoring-oriented radar intelligent classification and identification method.

[0029] The application has the following advantages and beneficial effects:

[0030] The application effectively overcomes the insufficient feature representation capability of the traditional method in low-altitude target track classification, and solves the performance loss problem of the deep model when processing unequal-length sequences under a specific platform.

[0031] The application proposes a comprehensive and efficient hybrid feature engineering scheme, which greatly improves the feature representation capability and classification accuracy. The multi-level feature set containing basic physical features and advanced derived features is creatively constructed. The feature set not only covers the instantaneous state information of target motion (such as coordinates, speed, and acceleration), but also describes the motion intention and aerodynamic characteristics of the target from multiple dimensions through a number of stability indicators (such as speed variance, height variation coefficient, etc.), maneuverability indicators (such as acceleration variance, trajectory curvature, maneuvering factor, etc.), and behavior pattern indicators (such as hovering index, climbing rate, etc.).

[0032] The feature engineering scheme provided by the application provides input data with extremely high information density and extremely strong discrimination for subsequent classification recognition models, fundamentally solves the problem of weak feature representation capability of traditional methods, lays a solid foundation for high-precision classification and recognition, and further significantly improves the classification accuracy.

[0033] The application proposes and effectively solves the problem that the model performance is impaired due to the lack of a mask mechanism in industrial platforms such as MATLAB. The introduced self-defined effective step extraction layer can receive the actual length information of the sequence, accurately shield the influence of the padding value in the forward propagation, and only output the results corresponding to the effective time steps; in the backward propagation, the gradient is only updated through the effective data. The self-defined layer effectively realizes the function equivalent to the mask mechanism in a deep learning environment (such as MATLAB) that does not support native masks, eliminates the pollution of padding noise on model training, enables the advanced CNN-LSTM hybrid model to run stably and efficiently on such platforms, and further ensures that the model can focus on learning the internal laws of effective data, thereby improving the generalization ability and classification accuracy of the model. Experimental verification shows that after using the self-defined effective step extraction layer of the application, the target recognition accuracy can be improved by about 3%.

[0034] The application realizes end-to-end high-precision and high-robustness target classification by fusing the synergistic advantages of hybrid features and adaptive networks. The organic combination of the above hybrid features and adaptive networks not only provides high-quality input for the network, but also ensures that such high-quality features are used most effectively through the adaptive network.

[0035] The complete classification scheme finally formed by the application has extremely high classification accuracy and robustness for track recognition of various targets (such as unmanned aerial vehicles, civil passenger aircraft, birds, etc.) in low-altitude environments. Experimental verification shows that the overall classification accuracy of the application scheme on a real scene test set reaches 93%. At the same time, the application has low hardware requirements, strong engineering practicability, and is easy to deploy and apply.

[0036] Since the application only relies on the track data output by the conventional detection of the radar system, without the aid of full-polarization radar, high-resolution radar images or complex micro-Doppler signal processing, the requirements for the radar hardware system are relaxed, the implementation cost is reduced, and the application scheme can be quickly deployed in existing radar systems, greatly promoting its practical application and industrialization in the low-altitude supervision field. In addition, in the unmanned aerial vehicle group cooperation scene or air defense system, the track is the core information shared across platforms, and the track-based recognition method facilitates direct interaction and data sharing between different sensors. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1An exemplary flow chart of a radar intelligent classification and identification method for low-altitude supervision provided by the present application is shown in FIG. 1.

[0038] Figure 2 An exemplary schematic diagram of a hybrid neural network model provided by the present application is shown in FIG. 2.

[0039] Figure 3 An actual measurement identification effect diagram of the radar intelligent classification and identification method and system for low-altitude supervision provided by the present application is shown in FIG. 3. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0041] Figure 1 An exemplary flow chart of a radar intelligent classification and identification method for low-altitude supervision provided by the present application is shown in FIG. 1. Figure 1 As shown in FIG. 1, the radar intelligent classification and identification method for low-altitude supervision provided by the present application includes the following contents:

[0042] Data preparation and labeling: first, the track data of various flight targets in the air is collected by the radar, then the real target type corresponding to each target track is identified by the photoelectric equipment, and then the real class label is added to each track, and finally the track data set is obtained. The real target type refers to the actual target type to which the target track data corresponds, including fixed-wing unmanned aerial vehicles, rotary-wing unmanned aerial vehicles, helicopters and birds, etc.

[0043] The track data set is processed to obtain track features; the track features include basic features and derived features, and the feature processing specifically includes data cleaning and feature extraction. The basic features include period, coordinate value, speed, acceleration, heading angle, track quality, and RCS, etc.; the derived features include speed variance, speed coefficient of variation, speed oscillation frequency, height variance, height coefficient of variation, height oscillation frequency, acceleration variance, acceleration coefficient of variation, trajectory curvature, climb rate, hovering index, and maneuver factor, etc. Data cleaning: first, identify potential problems in the track data (such as missing values, repeated values, outliers, etc.); for missing values, fill in with statistical values such as mean, median or mode; for repeated values, according to the business requirements, choose to retain unique values or merge processing; for outliers, after clarifying the cause, replace with statistical values. Feature extraction: extract conventional basic features (including the above period, coordinate value, speed, acceleration, heading angle, track quality, and RCS, etc.); extract the above derived features (including speed variance, speed coefficient of variation, and maneuver factor, etc.). Among them, the hovering index, the climb rate, and the trajectory curvature are innovative kinetic derived features, which have a significant effect in the classification identification system and can improve the identification accuracy and robustness.

[0044] The hovering index is a core indicator quantifying the target's ability to maintain position stability in three-dimensional space. The calculation process of the hovering index includes: first, calculating the proportion of track points with speed below a threshold value within the window ; second, calculating the sum of variances of the target's three-dimensional coordinates within the window ; and finally obtaining the hovering index . The calculation formula of the hovering index is:

[0045] ;

[0046] ;

[0047] ;

[0048] wherein, represents the hovering index; i represents the current time step; represents the proportion of track points with speed below a threshold value within the window; η represents a scaling coefficient; represents the sum of variances of the target's three-dimensional coordinates within the window; N is the length of the sliding window; k ∈ [i-N+1, i] represents the index of the time step within the sliding window; and represent the horizontal axis coordinates of the target track at the kth and k-1th time steps, respectively; and represent the vertical axis coordinates of the target track at the kth and k-1th time steps, respectively; and represents the vertical axis coordinate of the target track at the kth and k-1th time step, respectively; represents the resultant velocity of the track at the kth time step; represents the velocity threshold for determining the hovering of the track; represents the indicator function.

[0049] The climbing rate is used to measure the rate of change of the height of the target in the vertical direction, and the climbing rate is calculated by the following formula:

[0050] ;

[0051] ;

[0052] wherein, represents the climbing rate; i represents the current time step; represents the track velocity at the ith time step; sin represents the sine function; represents the climbing angle of the track at the ith time step; arctan represents the inverse tangent function; and represent the track height at the ith and i-1th time step, respectively; and represent the track horizontal distance at the ith and i-1th time step, respectively.

[0053] The track curvature is used to quantify the bending degree of the target track, and the instantaneous curvature is fitted by three consecutive sampling periods. Assuming that the track sequence points include three sequence points, i.e., the i-2th sequence point , the i-1th sequence point and the ith sequence point , the calculation formula of the track curvature is as follows:

[0054] ;

[0055] ;

[0056] ;

[0057] ;

[0058] wherein, represents the track curvature; i represents the current time step; , and represent the horizontal axis, vertical axis and vertical axis coordinates of the track at the ith time step, respectively; , and represent the horizontal axis, vertical axis and vertical axis coordinates of the track at the i-1th time step, respectively. , and denote the horizontal, vertical and vertical axis coordinates of the i-2th time step track respectively; = P i-2 - P i-1 denotes a forward vector, pointing from point P i-1 to P i-2 ; = P i - P i-1 denotes a backward vector, pointing from point P i-1 to P i ; = P i - P i-2 denotes a span vector, pointing from point P i-2 to P i ; denotes the length of the cross product of vectors and .

[0059] The track features are enhanced and compressed to obtain a mixed feature set, and based on the mixed feature set and the corresponding target track type, a training sample is constructed, including:

[0060] The track features are enhanced to obtain an enhanced feature dataset; the feature enhancement refers to expanding the feature dataset. For example, the feature dataset is expanded by rotating, translating and / or adding Gaussian noise, etc., to improve the robustness of the model.

[0061] The enhanced feature dataset is divided using stratified sampling to obtain a divided dataset; in each subset of the divided dataset, the proportion of each type of target remains consistent with the original track dataset. Stratified sampling is used for dataset division to ensure that the class proportion of each subset after division is consistent with the original data, ensuring the fairness of model evaluation, and thus building a model with good generalization ability for unknown data.

[0062] The features in the divided dataset are scaled using Z-Score standardization to obtain a scaled dataset; the data in the scaled dataset is scaled to a normal distribution with a mean of 0 and a variance of 1. Feature scaling can eliminate the volatility and dimension difference of feature data distribution, thereby improving the efficiency and performance of model training. To prevent feature data leakage, only the training set is used to calculate the parameters required for standardization (such as mean, standard deviation, etc.), and then the training set, validation set and test set are uniformly converted.

[0063] The feature importance evaluation is performed on the scaled data set, and the feature compression is performed according to the evaluation result, so as to obtain a compressed data set, and the compressed data set is used as a mixed feature set, including: the correlation between the features and the target variable is destroyed by using a random forest, the performance of a mixed neural network model is evaluated, and the performance degradation degree is obtained; the Lasso regression is used to force sparse solution, and the sparse compression of part of the features is compressed to zero, so as to obtain important features and unimportant features, and the compression degree is determined based on the values obtained after the important features and the unimportant features are compressed; the important features refer to the features that are not compressed to zero, and the unimportant features refer to the features that are compressed to zero; the features of a verification set are disturbed by using permutation importance on a classification model, the performance fluctuation of the classification model is output, the features are evaluated, and the fluctuation degree is obtained; the variance analysis is used to quantify the explanation ability of the features to the category difference, and the difference degree is obtained; the performance degradation degree, the compression degree, the fluctuation degree and the difference degree are weighted and comprehensively scored, and a comprehensive importance index is generated; and a set composed of multiple features with the comprehensive importance index greater than a preset dynamic threshold is used as the compressed data set. The purpose of feature compression is to remove noise and redundant features, and to improve the generalization ability of the model. The feature compression model is fitted based on the training set, and is applied to the feature processing of the verification set and the test set. For example, four kinds of feature importance evaluation methods are comprehensively used, which are a random forest algorithm, a Lasso regression, a permutation importance analysis and an ANOVA F value analysis. The random forest algorithm is used to evaluate the performance degradation degree of the model by destroying the correlation between the features and the target variable; the Lasso regression is used to compress the coefficients of the unimportant features to zero by using L1 regularization to force sparse solution; the permutation importance analysis is used to disturb the features of the verification set on the trained model, and the importance of the features is evaluated by the performance fluctuation of the model; and the ANOVA F value analysis is used to evaluate the importance of the features by quantifying the explanation of the features to the category difference.

[0064] The evaluation results of the above-mentioned methods are fused, the feature comprehensive importance index is generated by using a weighted comprehensive scoring method, the dynamic threshold is set based on the index, and only the feature subset with the feature comprehensive index higher than the threshold is retained, so that the compression of the feature space is realized.

[0065] The CNN+LSTM hybrid network architecture is used to construct a mixed neural network model with a self-defined effective step extraction layer; the self-defined effective step extraction layer filters the output sequence of the LSTM, and adaptively extracts effective features, so as to eliminate the model performance degradation problem caused by sequence padding noise. Figure 2As shown, the hybrid neural network model comprises a sequence input layer, a 1D convolution layer, a batch normalization layer, a first activation function layer, an LSTM layer, a self-defined effective step extraction layer, a first full connection layer, a second activation function layer, a second full connection layer, a Softmax layer and a classification output layer; the sequence input layer is used for inputting the hybrid features; the 1D convolution layer is used for convolution processing on the input hybrid features; the batch normalization layer is used for normalization processing on the hybrid features after convolution processing; the first activation function layer is used for activation processing on the hybrid features after normalization processing; the LSTM layer is used for long-short time feature extraction on the hybrid features after activation processing; the self-defined effective step extraction layer is used for effective feature extraction on the hybrid features after long-short time feature extraction and the hybrid features input by the sequence input layer, to obtain effective features; the first full connection layer is used for first full connection on the effective features; the second activation function layer is used for activation on the effective features after first full connection; the second full connection layer is used for second full connection on the effective features after activation; the Softmax layer is used for conversion on the effective features after second full connection, to obtain the probability distribution of target classification; and the classification output layer is used for outputting the target track classification result based on the probability distribution of target classification. Specifically, the CNN is used for feature extraction, to extract high-level and discriminative local features from the original sequence; then the feature sequence is sent to the LSTM layer, to learn the long-term dependence in the sequence by the LSTM layer. The LSTM can selectively remember or forget information through its gating mechanism (input gate, forget gate and output gate), and is suitable for sequence data processing; the LSTM layer is followed by the self-defined effective step extraction layer of the application, which has the core function of eliminating the padding noise introduced by the padding operation of unequal length sequences. The self-defined layer only transmits effective features from real data during forward propagation, and ensures that the learning signal only comes from effective data during backward propagation, so as to realize the equivalent model purification effect in the case of lacking native mask support in the framework, and further significantly improve the training stability and classification accuracy. After the output of the self-defined layer, the output vector is sent to the first full connection layer for transformation and combination, to further learn the nonlinear feature combination and lay the foundation for the final classification; then the features learned by the network are sent to the second full connection layer, to map the features to the "original score" corresponding to M target classes; and the "original score" is converted to a probability distribution by the softmax function. The final layer of the network is the output layer, which is used for calculating the loss function and completing model training or target prediction.

[0066] The hybrid neural network model is trained and verified by training samples to obtain a classification model, including: training on the training set, monitoring performance on the validation set, and optimizing and early stopping through hyperparameters until the performance of the hybrid neural network model on the validation set reaches the cycle termination condition (loss function converges or iteration number reaches threshold), to obtain the final hybrid neural network model. For example, the training and verification adopts a training method of cyclic iteration: model training → validation set evaluation → adjusting hyperparameters or model structure based on the verification result → training again. The cycle continues to be executed until the performance of the model on the validation set reaches the preset satisfactory standard. A test set independent of the training process (i.e. test samples not involved in any training link) is used to perform final evaluation on the final hybrid neural network model to obtain a generalization performance index; the hybrid neural network model whose generalization performance index meets the preset requirement is determined as the final classification model, otherwise the above training and verification cycle is returned to be executed again. The generalization performance preset requirement refers to the minimum generalization performance standard that a qualified classification model needs to reach.

[0067] The classification model is deployed to perform the same data processing procedure as described above on real-time track data collected by the radar, and the real-time hybrid features are input into the classification model to output target track classification results. The data processing procedure specifically includes: performing feature processing on real-time track data to obtain real-time track features; performing feature enhancement and compression processing on real-time track features to obtain real-time hybrid features. In addition, continuous optimization of the classification model is also achieved through the following steps: obtaining the real type of the target track; based on the real type and the target track classification result, iteratively optimizing the classification model. The specific deployment procedure is as follows: save the model with performance up to standard and the corresponding cleaning rules and feature extraction logic, and then deploy online; continuously monitor the model running performance after deployment, and collect new track data for model iterative optimization. For the newly collected data, perform the cleaning and feature processing procedure consistent with the model training phase; load the saved model in the deployment environment, input the preprocessed feature data into the model to perform prediction, output the final classification result and feed back to the subsequent decision system.

[0068] The application further provides a radar intelligent classification and identification system for low-altitude supervision, comprising a flight path data acquisition module, a flight path feature acquisition module, a mixed feature acquisition module, a model construction module, a model training module and an online processing module; the flight path data acquisition module is used for collecting flight path data of an aerial target by a radar to obtain a flight path data set; the flight path data set comprises target flight path data and target flight path types; the flight path feature acquisition module is used for performing feature processing on the flight path data set to obtain flight path features; the flight path features comprise basic features and derived features; the basic features comprise period, coordinate value, speed, acceleration, heading angle, flight path quality and RCS; the derived features comprise speed variance, speed variation coefficient, speed oscillation frequency, height variance, height variation coefficient, height oscillation frequency, acceleration variance, acceleration variation coefficient, trajectory curvature, climb rate, hovering index and maneuvering factor; the mixed feature acquisition module is used for performing enhancement and compression processing on the flight path features to obtain a mixed feature set, and constructing training samples based on the mixed feature set and the corresponding target flight path types; the model construction module is used for constructing a mixed neural network model with a self-defined effective step extraction layer by using a CNN-LSTM mixed network architecture; the self-defined effective step extraction layer is used for eliminating unequal-length sequence padding noise; the model training module is used for training and verifying the mixed neural network model by using the training samples to obtain a classification model; and the online processing module is used for deploying the classification model to perform data processing on real-time flight path data collected by the radar, obtaining real-time mixed features, inputting the real-time mixed features into the classification model, and obtaining target flight path classification results.

[0069] The radar intelligent classification and identification method for low-altitude supervision comprehensively utilizes the motion characteristics, trajectory morphology and geometric features of a target, and constructs a multi-dimensional mixed feature set. In addition to extracting basic physical features (such as coordinates, speed, acceleration, heading angle, etc.) representing the instantaneous state of the target directly observed or obtained by simple calculation, the application further constructs a series of derived features representing the deep law of the motion mode, specifically including: speed variance, speed variation coefficient, height variance, height variation coefficient, etc., used to determine whether the target is cruising at a constant speed or changing in motion; acceleration variance, maneuvering factor, trajectory curvature, etc., used to quantify the degree of severity of the target maneuver; hovering index, climb rate, etc., used to represent the behavior mode of the target; flight path quality, RCS change, etc., used to assist in determining the target type and credibility. The feature set is not simply a data stack, but forms a "feature language" for deep interpretation of the behavior of the flight path, providing extremely rich and high-quality input information for the subsequent deep learning model.

[0070] The application also designs a self-defined effective step extraction layer, which can efficiently and accurately process unequal length sequences in a deep learning environment without supporting a mask mechanism, and eliminate the negative impact of padding operation on model training. The solution to this problem includes: first, abandoning the built-in LSTM layer (such as lstmLayer in MATLAB), and adopting a completely self-defined LSTM layer; second, completely customizing the training loop, and separately processing each batch inside the loop to dynamically adapt to unequal length sequences.

[0071] The application adopts the scheme of standard LSTM combined with a self-defined effective step extraction layer, which is located after the LSTM sequence output. The self-defined layer receives the actual effective length of each sample while receiving the features of each time step of the LSTM output, and then extracts the feature output of the last effective time step of each sequence from all time steps of the LSTM output. In this method, the front-end calculation is completed by the highly optimized built-in LSTM, which not only has high GPU utilization and fast training speed, but also only needs to write a simple self-defined effective step extraction layer with clear logic, which has strong code robustness and is easy to debug, and can be seamlessly integrated into the built-in training loop.

[0072] As shown in FIG. 3, the radar intelligent classification and identification method and system for low-altitude supervision provided by the application can effectively identify the tracks of low-altitude targets and complete classification: the target batch number corresponding to point A is 22894, the target batch number corresponding to point B is 22962, the target batch number corresponding to point C is 23155, and the target batch number corresponding to point D is 22923. Among them, the targets with batch numbers 22894, 22923 and 22962 are identified as “flying birds”, which are verified by photoelectric equipment and the identification is accurate; the target with batch number 23155 is identified as “rotary wing unmanned aerial vehicle”, which is the “Ou 3E” unmanned aerial vehicle flown by our side, and the identification is accurate.

[0073] The above is only a preferred embodiment of the application and is not used to limit the application. For those skilled in the art, the application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A radar intelligent classification and identification method for low-altitude surveillance, characterized in that, include: Track data of aerial targets is collected by radar to obtain a track dataset; The track dataset includes target track data and target track types; Feature processing is performed on the track dataset to obtain track features; Track features include basic features and derived features; Basic characteristics include period, coordinates, velocity, acceleration, heading angle, track mass, and RCS; Derived characteristics include velocity variance, velocity coefficient of variation, velocity oscillation frequency, altitude variance, altitude coefficient of variation, altitude oscillation frequency, acceleration variance, acceleration coefficient of variation, trajectory curvature, rate of climb, hovering index, and maneuvering factor; the hovering index is an indicator that quantifies a target's ability to maintain positional stability in three-dimensional space, and the formula for calculating the hovering index is: ； ； ； in, Indicates the hovering index; i represents the current time step; This indicates the proportion of waypoints within the window whose speed is below a threshold. Indicates the scaling factor; The sum of variances of the three-dimensional coordinates of the target within the window; N is the length of the sliding window; k∈[i-N+1,i] represents the index of the time step within the sliding window; and These represent the x-axis coordinates of the target trajectory at the k-th and (k-1)-th time steps, respectively. and These represent the ordinates of the target trajectory at time steps k and (k-1), respectively. and These represent the vertical coordinates of the target trajectory at the k-th and (k-1)-th time steps, respectively. This represents the resultant velocity of the trajectory at the k-th time step; This indicates the speed threshold for determining whether a flight path is hovering; Indicates an indicator function; The track features are enhanced and compressed to obtain a hybrid feature set, and training samples are constructed based on the hybrid feature set and the corresponding target track type. A hybrid neural network model with a custom effective stride extraction layer is constructed using a CNN-LSTM hybrid network architecture. The custom effective stride extraction layer filters the output sequence of the LSTM and adaptively extracts effective features to eliminate the model performance degradation caused by sequence padding noise. The hybrid neural network model is trained and validated using training samples to obtain a classification model; A classification model is deployed to process real-time track data collected by radar, obtain real-time mixed features, and input the real-time mixed features into the classification model to obtain the target track classification result.

2. The radar intelligent classification and identification method for low-altitude surveillance according to claim 1, characterized in that, Rate of ascent measures the rate of change of a target's altitude in the vertical direction. The formula for calculating the rate of ascent is: ； ； in, Indicates the climb rate; i represents the current time step; Let represent the resultant velocity of the trajectory at the i-th time step; sin represents the sine function. represents the climb angle of the trajectory at the i-th time step; arctan represents the arctangent function; and These represent the altitudes of the track at time steps i and i-1, respectively. and represents the horizontal distance of the track at the i-th and i-1th time steps, respectively.

3. The radar intelligent classification and identification method for low-altitude surveillance according to claim 1, characterized in that, Trajectory curvature is used to quantify the degree of curvature of a target trajectory. The formula for calculating trajectory curvature is: ； ； ； ； in, Indicates the trajectory curvature; i represents the current time step; , and These represent the horizontal, vertical, and y coordinates of the trajectory at the i-th time step, respectively. , and These represent the horizontal, vertical, and y-axis coordinates of the trajectory at the (i-1)th time step, respectively. , and These represent the horizontal, vertical, and y-axis coordinates of the trajectory at the (i-2)th time step, respectively. Represents the forward vector; Represents the backward vector; Represents the span vector; Representing vectors and The modulus of the cross product.

4. The radar intelligent classification and identification method for low-altitude surveillance according to claim 1, characterized in that, The track features are enhanced and compressed to obtain a hybrid feature set, including: Feature enhancement processing is performed on the track features to obtain an enhanced feature dataset; feature enhancement processing refers to expanding the feature dataset. Stratified sampling is performed on the enhanced feature dataset to obtain a partitioned dataset; the target proportion of each category in the subset of the partitioned dataset is consistent with that of the track dataset. Z-Score standardization is used to scale the features in the split dataset to obtain a scaled dataset; the data in the scaled dataset are scaled to a normal distribution with a mean of 0 and a variance of 1. By evaluating the feature importance of the scaled dataset and compressing the features based on the evaluation results, a compressed dataset is obtained, which is then used as a hybrid feature set.

5. The radar intelligent classification and identification method for low-altitude surveillance according to claim 4, characterized in that, By evaluating the feature importance of the scaled dataset and compressing the features based on the evaluation results, a compressed dataset is obtained, including: The performance of the hybrid neural network model is evaluated by examining the correlation between the destruction features of random forests and the target variable, and the degree of performance degradation is obtained. By using Lasso regression sparse solutions, some features are sparsely compressed to zero to obtain important and unimportant features. The degree of compression is determined based on the values ​​obtained after compression of important and unimportant features. Important features are those that are not compressed to zero, and unimportant features are those that are compressed to zero. By permuting importance, the features of the validation set are permuted and disturbed in the classification model. The volatility evaluation features of the classification model are then used to obtain the degree of volatility. Analysis of variance is used to quantify the explanatory power of features for class differences, thereby obtaining the degree of difference; A comprehensive importance index is generated by weighting and comprehensively scoring the degree of performance degradation, compression, volatility, and variability. The set of multiple features whose comprehensive importance index is greater than the preset dynamic threshold is used as the compressed dataset.

6. The radar intelligent classification and identification method for low-altitude surveillance according to claim 1, characterized in that, The hybrid neural network model includes a sequence input layer, a 1D convolutional layer, a batch normalization layer, a first activation function layer, an LSTM layer, a custom effective stride extraction layer, a first fully connected layer, a second activation function layer, a second fully connected layer, a Softmax layer, and a classification output layer. The sequence input layer is used to input mixed features; 1D convolutional layers are used to perform convolutional processing on the mixed features of the input; Batch normalization layers are used to normalize the mixed features after convolution; The first activation function layer is used to activate the normalized mixed features; LSTM layers are used to extract long and short time features from the blended features after activation processing; A custom effective step size extraction layer is used to extract effective features from the mixed features extracted from long and short time features and the mixed features from the input sequence input layer, thus obtaining effective features. The first fully connected layer is used to perform the first full connection on the effective features; The second activation function layer is used to activate the effective features after the first fully connected layer; The second fully connected layer is used to perform a second full connection on the activated valid features; The Softmax layer is used to transform the effective features after the second fully connected layer to obtain the probability distribution of the target classification. The classification output layer is used to output the target track classification result based on the probability distribution of target classification.

7. The radar intelligent classification and identification method for low-altitude surveillance according to claim 1, characterized in that, By training and validating the hybrid neural network model using training samples, a classification model is obtained, including: The hybrid neural network model is trained on the training set, and its performance is monitored on the validation set. It is then tuned and early stopped using hyperparameters until the performance of the hybrid neural network model on the validation set reaches the loop termination condition, thus obtaining the final hybrid neural network model. The final hybrid neural network model is evaluated using a test set to obtain generalization performance metrics; the test set refers to independent training samples that have not participated in any training process. The final hybrid neural network model that achieves the required generalization performance index is used as the classification model.

8. The radar intelligent classification and identification method for low-altitude surveillance according to claim 1, characterized in that, Data processing is performed on the real-time track data acquired by radar to obtain real-time hybrid features, including: Feature processing is performed on real-time track data to obtain real-time track features; The real-time track features are enhanced and compressed to obtain real-time hybrid features; Also includes: Obtain the actual track type of the target track; The classification model is continuously optimized based on the actual track type and target track classification results.

9. A radar intelligent classification and identification system for low-altitude surveillance, utilizing the radar intelligent classification and identification method for low-altitude surveillance as described in any one of claims 1-8, characterized in that, It includes a trajectory data acquisition module, a trajectory feature acquisition module, a hybrid feature acquisition module, a model building module, a model training module, and an online processing module; The trajectory data acquisition module is used to collect trajectory data of aerial targets through radar to obtain a trajectory dataset; the trajectory dataset includes target trajectory data and target trajectory type; The track feature acquisition module is used to perform feature processing on the track dataset to obtain track features; Track features include basic features and derived features; Basic characteristics include period, coordinates, velocity, acceleration, heading angle, track mass, and RCS; Derived features include velocity variance, velocity coefficient of variation, velocity oscillation frequency, altitude variance, altitude coefficient of variation, altitude oscillation frequency, acceleration variance, acceleration coefficient of variation, trajectory curvature, rate of climb, hovering index, and maneuvering factor; The hybrid feature acquisition module is used to enhance and compress track features to obtain a hybrid feature set, and to construct training samples based on the hybrid feature set and the corresponding target track type. A hybrid neural network model with a custom effective stride extraction layer is constructed using a CNN-LSTM hybrid network architecture. The custom effective stride extraction layer filters the output sequence of the LSTM and adaptively extracts effective features to eliminate the model performance degradation caused by sequence padding noise. The model training module is used to train and validate the hybrid neural network model using training samples to obtain a classification model; The online processing module is used to deploy the classification model, process the real-time track data collected by radar to obtain real-time mixed features, and input the real-time mixed features into the classification model to obtain the target track classification result.

Images