Radar low-slow small target classification identification method based on probability oversampling

Abstract

The invention belongs to the technical field of radar data processing, and provides a radar low-slow small target classification and identification method based on probability oversampling. According to the method, radar clutters and trace point information of low-slow small targets are collected and processed, for single-frame radar data, the clutter data volume is significantly larger than the data volume of the low-slow small targets, and the whole data set is an unbalanced data set. Clutter data of original unbalanced data is clustered through a clustering method, various clustered sub-data sets and minority data are combined into unbalanced sub-data sets, each unbalanced data subset is balanced through a probability oversampling method, a model is established to obtain a decision matrix, and a better classification result is obtained through majority voting. By adopting the method, the defects of low classification accuracy and poor clutter filtering effect of the low, slow and small radar target are overcome, additional hardware does not need to be added, classification and identification of the radar clutter and the low, slow and small target are realized in a single radar antenna period, and the radar clutter is better filtered in a data processing level.

Description

Technical Field

[0001] The invention belongs to the technical field of radar data processing. Background Art

[0002] In an unbalanced data set, the number of samples of some classes is significantly greater than that of other classes. In classification problems, the class with more data is often called the majority class data, and the class with less data is called the minority class data. These problems have a common feature, that is, the information of the minority class samples is the focus of attention. For example, in the problem of credit card transaction fraud monitoring, fraudulent users are only a minority, but they are the focus of researchers; in the diagnosis process of cancer patients, cancer patients are a minority and are the focus of researchers. If these minority events cannot be identified normally, serious consequences will occur. Classical classification methods cannot be directly applied to unbalanced data sets. Due to the imbalance of data, the model learned from the majority class data is prone to overfitting, while the minority class data lacks sufficient samples, and the model's recognition ability for the minority class data is limited, which easily leads to underfitting of the model. Therefore, the research on unbalanced data classification has broad application prospects and practical significance.

[0003] For radar, in a complex electromagnetic environment or a strong clutter environment, there will be quite a lot of irregular clutter areas within the radar coverage area. These may be static clutter such as ground buildings, hillsides, and electric poles that are stationary relative to the radar, or dynamic clutter such as clouds, rain, waves, and trees swaying in the wind that are moving relative to the radar. The radar data collected by the low, slow, and small detection radar includes both clutter data and low, slow, and small target data. The amount of clutter data in each frame of radar data is significantly greater than the amount of low, slow, and small target data, resulting in the entire data set being an unbalanced data set. At the same time, low, slow, and small targets have the characteristics of low flight altitude, slow flight speed, low target radar echo intensity, easy to be blocked by the surrounding environment, and the target is easily submerged in environmental clutter. Therefore, the classification and identification of low, slow, and small targets has important practical significance for national defense security. Summary of the invention

[0004] In order to overcome the problems of difficulty in identifying low, slow and small radar targets and low recognition rate at the traditional data processing level, the present invention provides a radar low, slow and small target classification and recognition method based on probability oversampling, which is real-time, universal and efficient.

[0005] In order to achieve the above technical objectives, the present invention proposes a radar low, slow and small target classification and recognition method based on probability oversampling, which comprises the following steps:

[0006] Step (1): Collect clutter and point trace data of radar detection targets as training samples, and confirm the true class label of the radar target through optoelectronic equipment;

[0007] Step (2): preprocess the data in step (1) to eliminate the impact caused by different dimensions of the data;

[0008] Step (3): Cluster the clutter data in step (2), select a reasonable number of clusters K, and generate an unbalanced data subset L;

[0009] Step (4): Balance the data set in step (3) by using the probability oversampling method, establish a model for classification, and obtain a decision matrix for the test data;

[0010] Step (5): The data in step (4) is voted by majority class to determine the final class label of the test data, thereby obtaining the radar low, slow and small target data classification result.

[0011] The clutter and target point trace data features described in step (1) mainly include: period, time, point trace azimuth, point trace distance, point trace elevation angle, Doppler velocity, point trace amplitude, elevation angle quality and other feature dimensions. The original data set is: Where n and m are the number of data and the number of feature attributes respectively.

[0012] Furthermore, the preprocessing of the point trace data in step (2) may be a minimum-maximum normalization method, including: where x n is the standardized value, x min and x max are the minimum and maximum values ​​of different attribute values ​​of the data set respectively. The above preprocessing of point trace data eliminates the influence caused by different dimensions of data.

[0013] Furthermore, step (3) divides the original data into majority class data and minority class data, and processes them separately. The majority class data is clustered into specific classes using the commonly used and easy-to-operate k-means clustering algorithm to obtain a majority class sub-dataset C. Each majority class sub-dataset obtained is merged with the minority class data to obtain a sub-dataset L. The first stage of the algorithm determines a reasonable number of clusters K. First, the number of clusters is set to be more than two classes, otherwise there is no clustering process. Initialize the maximum number of clusters K max And the reasonable number of clusters Since the Kappa value is a good indicator of the credibility of the classification model, the Kappa statistic is preferably used to help set a reasonable number of clusters, and the Kappa value of the subclassifier is used to reversely determine the reasonable number of clusters that the data set should be classified into. In order to ensure the reliability of the classification model, the present invention uses the degree of Kappa value higher than slightly (i.e., Kappa value greater than 0.2) as a standard performance indicator. First, the majority class data of the data set is divided into groups according to the number of clusters from 2 to K. maxCluster the data in turn, merge the obtained data of each category with the original minority category data to obtain the unbalanced data subset L, build a model for L through the classifier, and obtain the Kappa value L of each subset of the L data set Kappa , if the minimum value of all Kappa values ​​is greater than 0.2, that is The number of clusters corresponding to the initial selection is taken as the number of reasonable clusters to be selected, that is, K set =K set ∪k. Kappa value greater than 0.2 means that the model established by this clustering method is credible. This process is iterated to obtain a reasonable cluster number set K set , set the minimum number of clusters as the reasonable number of clusters, that is, K = min (K set ), which facilitates the fast and efficient execution of the algorithm.

[0014] Furthermore, step (4) uses the fast convergence Gibbs probability oversampling method to synthesize new minority class data to balance the unbalanced sub-dataset L, and then obtain the balanced sub-dataset S, achieving the overall balance of the data set, and then build a model on the balanced sub-dataset to obtain the decision matrix. First, the Chow-Liu algorithm is used to approximate the probability distribution of the minority class data in each subset in L, and new data is extracted from the approximate probability distribution through Gibbs sampling to balance the sub-dataset. Demonstrate the ability to synthesize new data by sampling from an approximate univariate conditional distribution In general, the initial values ​​of random variables for Gibbs sampling are randomly selected from the state space of the attribute. Fast-converging Gibbs sampling takes the minority class data points as a set of initial samples, and then performs Gibbs sampling for each minority class data. It generates multiple Markov chains, each chain starts with a different minority class sample, unlike traditional Gibbs sampling that generates a very long Markov chain. Its initial values ​​are directly selected from the minority class samples, achieving faster convergence when generating new data. The Gibbs method is a Markov chain Monte Carlo method. It creates a Markov chain of random variables that converge to the target probability distribution. At each sampling step, the Gibbs method considers the univariate conditional distribution of each dimension. The value of each dimension depends on the values ​​of other dimensions and the previous values ​​of the same dimension, and new samples are generated using the univariate conditional distribution of each dimension. Gibbs sampling depends on two important factors, one is the number of iterations to generate samples in order to achieve a stable distribution, and the other is the number of consecutive samples discarded from the Markov chain. The newly generated minority class data considers the original probability distribution of the data, has good representativeness, and eliminates the data skewness to obtain a balanced sub-dataset S. Then, the model M=M is ​​established for the balanced sub-dataset S through the base classifier. k , the decision matrix D = M about the test data is obtained by classifying the test data through the established model k (tm ).

[0015] Furthermore, step (5) determines the final class label of the test data by majority voting. By clustering the majority class data, the overall imbalance rate of the original data set is reduced, and a better model is established for it. The sub-model established on the balanced sub-data set not only has a strong classification ability for the majority class data (negative class data) of specific features, but also has a strong classification performance for the positive class data. Therefore, the method proposed in the present invention ensures the reliability of model classification through a relatively strict voting method. That is, for a single test data, the classification results of each sub-model of model M are all positive, and its label is finally set to the positive class, that is, Otherwise, it is set as a negative class. This design method considers the classification results of all classification models on the same test data, ensuring that the minority class data is correctly classified by all classification models, and the obtained data class labels have strong credibility.

[0016] The beneficial effects of the present invention include: using a clustering method to process the original unbalanced data set reduces the overall imbalance rate of the original data set, and obtains an unbalanced data subset with a smaller imbalance rate. The probability distribution of minority class data is approximated by a probability method, and minority class data with a real probability distribution is sampled and synthesized, so that the newly synthesized data has good representativeness, and the classification performance of the original unbalanced data set is better improved, and the classification of clutter and slow and small targets is achieved, thereby suppressing the onset of clutter and obtaining a better clutter filtering effect. The present invention has a high clutter classification accuracy and a high low and slow small target classification and recognition capability, does not require the addition of additional hardware, has a low implementation cost, can be promoted and applied to related product fields, and has a good prospect for promotion and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Flow chart of an embodiment of the present invention DETAILED DESCRIPTION

[0018] Figure 1 For the implementation process of the present invention, the present invention is illustrated below with reference to examples.

[0019] 1. Collect clutter and low-speed small target trace information through Ku-band low-speed small target detection radar. Low-speed small target data includes target trace information such as drones (DJI Phantom 4 Pro) and birds. Confirm the real class label of low-speed small targets through optoelectronic equipment. The target trace data features mainly include: period, time, trace azimuth, trace distance, trace elevation, Doppler velocity, trace amplitude, elevation quality and other feature dimensions.

[0020] 2. Data preprocessing: In order to make the attributes of the data input to the model not affected by the dimension, all input data are standardized to the range of 0-1 using the minimum-maximum standardization method.

[0021] 3. Kappa value is an effective indicator of test data continuity and classification model credibility, and evaluates whether the model performance is within an acceptable range. When the classifier deals with unbalanced data problems, there is a phenomenon of high accuracy and low Kappa value. Its value is between -1 and 1. The part less than 0 is poor, indicating that the model is below the acceptable level. The part greater than 0 is evenly divided into 5 levels (slight, fair, moderate, substantial, perfect) to reflect the credibility level of the classification model. Kappa is defined as Where P0 represents the percentage of correct classification, P c Indicates the possibility of correct classification. In the embodiment of the present invention, the degree to which the Kappa value is higher than slightly (ie, the Kappa value is greater than 0.2) is used as a standard performance indicator.

[0022] 4. Approximate the probability distribution of minority class data in each subset in L through the fast convergence Gibbs sampling method, and extract new data from the approximate probability distribution through Gibbs sampling to balance the sub-datasets. Demonstrate the ability to synthesize new data by sampling from an approximate univariate conditional distribution Then, the model M=M is ​​established for the balanced sub-dataset S through the base classifier k , the decision matrix D = M about the test data is obtained by classifying the test data through the established model k (t m ).

[0023] 5. By clustering the majority class data, the overall imbalance rate of the original data set is reduced, and a better model is established for it. The sub-model established on the balanced sub-dataset has a strong classification ability not only for the majority class data (negative class data) of specific features, but also for the positive class data. That is, for a single test data, each sub-model of model M will set its label to the positive class only if all its classification results are positive, otherwise it will be set to the negative class. This design method considers the classification results of all classification models for the same test data, ensures that the minority class data is correctly classified by all classification models, and then filters out and suppresses the onset of clutter.

[0024] In order to verify the effectiveness of the proposed method, the low-speed and slow-speed small target point data used in the experiment comes from the data collected by a certain type of anti-UAV Ku radar, and the collection location is a certain land test site. The point data set includes clutter and low-speed and slow-speed small target point information. The low-speed and slow-speed small target data includes target point information such as drones (DJI Phantom 4 Pro) and flying birds. The collected data are all confirmed by optoelectronic equipment to identify their target categories. Clutter and low-speed and slow-speed small target types are represented by category 1 and category 2, respectively.

[0025] The following table shows the radar trace data of 6 different time periods, and the classification and recognition results of clutter and slow and small targets are respectively performed. The cooperative target flying during time periods 1 and 2 is the DJI Phantom 4 Pro. From the table below, it can be seen that the accuracy of clutter classification and recognition in time periods 1 and 2 is 0.946 and 0.924, and the accuracy of slow and small target classification and recognition is 0.950 and 0.962 respectively. The cooperative targets flying during time periods 3 and 4 are the DJI Phantom 4 Pro and the bird. From the table below, it can be seen that the accuracy of clutter classification and recognition in time periods 3 and 4 is 0.953 and 0.941, and the accuracy of slow and small target classification and recognition is 0.916 and 0.905 respectively. The target flying during time periods 5 and 6 is the bird. From the table below, it can be seen that the accuracy of clutter classification and recognition in time periods 5 and 6 is 0.927 and 0.939, and the accuracy of slow and small target classification and recognition is 0.868 and 0.892 respectively. Experiments show that the method proposed in the present invention significantly improves the accuracy of clutter and low, slow and small target classification, improves the radar clutter filtering effect, and enhances the radar low, slow and small target detection capability, providing a guarantee for better automatic starting in subsequent radar data processing.

[0026]

[0027] The present invention provides a radar low-speed and slow small target classification and recognition method based on probability oversampling. The application shows that the present invention achieves higher clutter classification accuracy and higher low-speed and slow small target classification and recognition capability without adding additional hardware and low implementation cost. It can be widely used in related product fields and has a good prospect for promotion and application.

Claims

1. A radar low, slow and small target classification and recognition method based on probability oversampling, characterized by: Step (1): Collect clutter and point trace data of radar detection targets as training samples, and confirm the true class label of the radar target through optoelectronic equipment; Step (2): preprocess the data in step (1) to eliminate the impact caused by different dimensions of the data; Step (3): Cluster the clutter data in step (2), select a reasonable number of clusters K, and generate an unbalanced data subset L; Step (4): Balance the data set in step (3) by using the probability oversampling method, establish a model for classification, and obtain a decision matrix for the test data; Step (5): The data in step (4) is voted by majority class to determine the final class label of the test data, thereby obtaining the radar low, slow and small target data classification result.

2. According to the probability oversampling based radar low, slow and small target classification and recognition method according to claim 1, it is characterized by: The preprocessing of the point trace data in step (2) is a minimum-maximum standardization method, including: The original data set X is: Normalized value x n for: Where n and m are the number of data and the number of feature attributes respectively, min and x max They are the minimum and maximum values ​​of different attribute values ​​in the data set.

3. The method for classifying and identifying radar low, slow and small targets based on probability oversampling according to claim 1 is characterized by: The process of clustering the clutter data in step (3) includes: dividing the original clutter data into majority class data and minority class data, and processing them separately; clustering the majority class data into specific classes through a clustering algorithm to obtain a majority class sub-dataset C, merging each obtained majority class sub-dataset with the minority class data to obtain a sub-dataset L; determining the number of clusters K, and initializing the maximum number of clusters K max And the reasonable number of clusters The Kappa value is slightly higher than the standard performance indicator: First, the majority class data of the dataset is divided into clusters from 2 to K according to the number of clusters. max Cluster the data in turn, merge the obtained data of each category with the original minority category data to obtain the unbalanced data subset L, build a model for L through the classifier, and obtain the Kappa value L of each subset of the L data set Kappa , if the minimum value of all Kappa values ​​is greater than 0.2, that is The number of clusters corresponding to the initial selection is taken as the number of reasonable clusters to be selected, that is, K set =K set ∪k, this process is iterated to obtain a reasonable clustering number set K set , set the minimum number of clusters as the reasonable number of clusters, that is, k = min (K set ).

4. The method for classifying and identifying low, slow and small radar targets based on probability oversampling according to claim 1 is characterized in that: The probability oversampling method in step (4) includes: synthesizing new minority class data to balance the unbalanced sub-dataset L, thereby obtaining a balanced sub-dataset S, achieving the overall balance of the data set, and then building a model on the balanced sub-dataset to obtain a decision matrix; using the Chow-Liu algorithm to approximate the probability distribution of the minority class data of each subset in L, and extracting new data from the approximate probability distribution through Gibbs sampling to balance the sub-datasets. Demonstrate the ability to synthesize new data by sampling from an approximate univariate conditional distribution The newly generated minority class data considers the original probability distribution of the data. The new data eliminates the skewness of the data set and obtains a balanced sub-dataset S. Then, the model M=M is ​​established for the balanced sub-dataset S through the base classifier. k , the decision matrix D=M about the test data is obtained through the established model classification k (t m ).

5. The method for classifying and identifying radar low, slow and small targets based on probability oversampling according to claim 1 is characterized by: The step (5) of determining the final class label of the test data by majority class voting includes: for a single test data, each sub-model of the model M performs the classification of the data t m The classification results are all positive, and then the label is finally set to the positive class, that is, Otherwise it is set as negative class.

Images