Radar target recognition method and device for mislabeling and imbalanced samples

Abstract

The application discloses a radar target recognition method and device for mislabeling and unbalanced samples, and the method comprises the following steps: pre-processing target data collected by a radar to obtain all training data, and dividing the training data into a training set and a test set; constructing a classification model, performing primary training on the training set, determining the optimal iteration number and a loss function; retraining the constructed classification model, setting the iteration number, calculating the loss function mean value of each sample of a certain type of sample, and removing the mislabeled samples; modifying the loss function of the primary training classification model, and performing secondary training by using the removed training data; and using the secondary trained classification model to perform real-time recognition on the radar target. The application can identify and remove mislabeled samples of the radar, eliminate the huge negative influence of the mislabeling on the recognition task, process unbalanced samples, and improve the recognition effect on the radar target.

Description

Technical Field

[0001] This invention relates to the field of radar signal feature extraction and target recognition technology, specifically to a radar target recognition method and apparatus for mislabeled and imbalanced samples. Background Technology

[0002] With the development of artificial intelligence technology, research has begun on the intelligent transformation of various equipment, including radar. Using radar data for target classification and identification has become a hot topic. Relevant research materials include Chinese patents CN201610896005.6, CN202011419393.1, CN201810851549.X, CN202111110784.X, and BK Kim and HS Kang et al.'s "Drone Classification Using Convolutional Neural Networks With Merged Doppler Images". "Classification of Birds and UAVs based on RadarPolarimetry" by Karl Erik Olsen and Hugh Griffiths, etc.

[0003] The mainstream research method for this topic uses supervised methods, which require collecting and labeling historical data on various targets such as drones, fixed-wing aircraft, and birds. However, due to the presence of non-cooperative targets, manual labeling may be incorrect, such as labeling man-made flying objects as birds. For example, when collecting drone data, in addition to drones launched by our own side, drones launched by others, i.e., non-cooperative targets, may appear within the radar detection range, and these non-cooperative targets are often misidentified as birds by humans.

[0004] Current patent literature has not yet discussed this issue; in order to eliminate the impact of mislabeling on the recognition effect, and considering that the amount of radar target data varies greatly in practical applications, this invention proposes a radar target recognition method and device for mislabeled and imbalanced samples. Summary of the Invention

[0005] The radar target recognition method and device are provided to overcome the deficiencies in the prior art, and the method comprises the following steps: firstly, dividing labeled radar target data into a training set and a test set, training a deep neural network classification model by using the training set, and determining the iteration number of the model according to the performance on the test set; then, retraining the classification model by using all data, calculating the loss of the model by inputting the bird sample, finding out the mislabeled sample and eliminating it; finally, retraining the classification model by using the data with the error label removed to perform radar target recognition.

[0006] The technical scheme of the present application is implemented as follows:

[0007] A radar target recognition method for mislabeled and imbalanced samples comprises the following steps:

[0008] Step a), pre-processing the labeled target data collected by the radar to obtain all training data T, and dividing the all training data T into a training set Tr and a test set Ts;

[0009] Step b), constructing a classification model, training the classification model by using the training set Tr in step a), and determining the optimal iteration number E of the classification model according to the performance of the classification model on the test set Ts p ; the loss function of the classification model is

[0010] wherein C is the number of sample categories, p = [p1, …, pi, …, pC] is a probability distribution, each element pi represents the probability that the sample belongs to the i-th category; y = [y1, …, yi, …, yC] is the One-Hot representation of the sample label, yi = 1 when the sample belongs to the i-th category, otherwise yi = 0; and a is the weighted coefficient of the i-th category, which is inversely proportional to the number of samples of the i-th category. C i C i i i

[0011] Step c), retraining the classification model constructed in step b) by using all the training data T in step a), and setting the iteration number of the training to E o ; E o is a positive integer not exceeding the optimal iteration number E p determined in step b);

[0012] ​​​​​​Step d), inputting each class of samples in the total training data T of step a) into the classification model retrained in step c) respectively, calculating the loss function mean value of each sample of the class; the samples corresponding to the maximum K loss function mean values are the mislabeled samples in the class, and the mislabeled samples are all removed from the total training data T of step a); wherein K is a predetermined positive integer; step d) is performed for each class of samples to remove all mislabeled samples in all classes of samples;

[0013] Step e), changing the loss function of the classification model trained in step b) to Training the training data T from which the mislabeled samples are removed in step d) again; γ in the loss function is a predetermined positive number; using the retrained classification model to perform real-time identification on the radar target.

[0014] As a further optimization, in step a), the target data at least includes high-dimensional time series data composed of target track, speed and radar scattering cross section information; the preprocessing includes interpolation and cleaning of the data, and cutting the high-dimensional time series data into equal-length standard sub-data sequences.

[0015] As a further optimization, in step b), the classification model constructed is a deep neural network classification model, and the deep neural network is one of CNN structure, RNN structure, LSTM structure and Transformer structure.

[0016] As a further optimization, in step b), the optimal iteration number E is determined according to the performance of the classification model on the test set Ts p , and the specific method is: calculating the auc value of the roc curve of the classification model on the test set Ts at different iteration numbers, and the iteration number corresponding to the maximum auc value is the optimal iteration number E p .

[0017] As a further optimization, in step d), each class of samples in the total training data T of step a) is input into the classification model trained in step c) respectively, and the loss function mean value of each sample of the class is calculated, and the specific method is:

[0018] d.1), the standard sub-data sequence obtained by cutting each sample in a certain class is input into the classification model trained in step c) respectively;

[0019] d.2), the loss function values of all standard sub-data sequences of the same flying bird sample ω j in the same class are calculated , wherein 0<j≤J, J is the total number of samples of the category in the training data T described in step a), is the sample of the category j The number of standard sub-data sequences cut out;

[0020] d.3) The average loss function value of each sample in the category is obtained:

[0021] As a further optimization, in step d), the sample corresponding to the maximum average loss function value is the mislabeled sample in the category, and all mislabeled samples are removed from the training data T described in step a). The specific method is:

[0022] d.4), E(L(ω j )) is arranged from large to small, and the samples corresponding to the top K average loss function values are mislabeled samples;

[0023] d.5) All standard sub-data sequences cut out from the mislabeled samples are removed from the training data T described in step a).

[0024] As a further optimization, in step e), the radar target is identified in real time, and the specific method is:

[0025] The time series data of the target detected by the radar is preprocessed according to step a), is continuously cut into standard sub-data sequences of equal length, and is sequentially sent into the classification model obtained in step e) to calculate the probability of belonging to each category. The probability of the target detected by the radar belonging to a category is obtained by averaging the probabilities of the sub-data sequences cut out from the target in the category. The category with the highest probability is the identification result of the target.

[0026] Meanwhile, the technical scheme of the present application also includes:

[0027] A radar target identification device for mislabeled and unbalanced samples, comprising a radar, a computer readable storage medium, a processor, a communication interface and a communication bus, wherein:

[0028] The radar is used to detect targets and collect labeled target data.

[0029] The communication bus communicates between the radar, the computer readable storage medium and the processor through the communication interface.

[0030] The computer readable storage medium is used to store instructions.

[0031] The processor is used to call instructions to execute the radar target identification method for mislabeled and unbalanced samples as described above.

[0032] The beneficial effects of the present application are as follows by adopting the technical scheme:

[0033] The present application eliminates the huge negative influence of the wrong labeling on the identification task by identifying and removing the wrong labeled samples of the radar, and improves the identification effect of the radar target by processing the unbalanced samples. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows.

[0035] Figure 1 The flow chart of the radar target identification method for mislabeling and unbalanced samples of the present application.

[0036] Figure 2 The deep neural network structure diagram of the embodiment of the present application is given.

[0037] Figure 3 The structure diagram of the radar target identification device for mislabeling and unbalanced samples of the present application. DETAILED DESCRIPTION

[0038] The technical solutions of the embodiments of the present application will be explained and described in combination with the drawings of the embodiments of the present application. However, the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0039] Embodiment one;

[0040] Referring to the drawings Figures 1-2 A radar target identification method for mislabeling and unbalanced samples, specifically comprising the following steps:

[0041] a). Preprocessing the labeled target data collected by the radar to obtain all training data T, and dividing T into a training set Tr and a test set Ts;

[0042] b). Constructing a deep neural network classification model, initially training the classification model with the training set Tr of step a), and determining the optimal iteration number E according to the performance of the classification model on the test set Ts p ;

[0043] The loss function of the classification model is Wherein, C is the number of sample categories, p=[p1,……,p C ] is a probability distribution, and each element p i represents the probability that the sample belongs to the i-th category; y=[y1,……,y Cis the One-Hot representation of sample label, y i = 1 when the sample belongs to the i-th class, otherwise y i = 0; a i is the weighting coefficient of the i-th class, which is inversely proportional to the number of samples in the i-th class;

[0044] c). Retrain the classification model constructed in step b) with all the training data T in step a), and set the number of iterations E o ; E o is a positive integer not exceeding the optimal number of iterations E p in step b);

[0045] d). Input a certain class of samples in the training data T in step a) into the classification model trained in step c) respectively, and calculate the average loss function value of each sample in this class; the samples corresponding to the K largest average loss function values are the mislabeled samples in this class, which are all removed from the training data T in step a); K is a predetermined positive integer; step d) is performed for each class of samples to remove all mislabeled samples in all classes;

[0046] e). Change the loss function of the classification model in step b) to Use the training data T with the removed mislabeled samples in step d) to train again; γ in the loss function is a predetermined positive number, and the meanings of other variables are the same as in step b); the classification model trained in step e) is used for real-time identification of radar targets.

[0047] The identification method identifies and removes mislabeled samples of the radar, eliminating the huge negative impact of mislabeling on the identification task; and by processing unbalanced samples, the identification effect of the radar target is improved.

[0048] Preferably, the target data in step a) includes high-dimensional time series data composed of target track, speed, and radar scattering cross-section information; the preprocessing in step a) includes interpolation and cleaning of data, and cutting the high-dimensional time series data into equal-length standard sub-data sequences.

[0049] In this embodiment, the categories of the labeled target data collected by the radar include unmanned aerial vehicles, birds, fixed-wing training aircraft, etc.

[0050] Preferably, the deep neural network in step b) can be a CNN structure, or an RNN structure, or an LSTM structure, or a Transformer structure; the optimal number of iterations E pThe AUC of the ROC curve on the test set is calculated for different iteration numbers of the classification model, and the iteration number corresponding to the maximum AUC is the optimal iteration number E p ;

[0051] The loss function set in step b considers the influence of unbalanced samples, and the influence is eliminated by inversely weighting the sample quantity;

[0052] In this embodiment, a one-dimensional CNN structure is selected, as shown in Figure 2 Conv[2,1,6], padding=same means that the convolution kernel length is 2, the step is 1, the channel number is 6, and padding=same means that padding is performed to make the input and output lengths the same.

[0053] In this embodiment, E o =[0.7·E p ] in step c);

[0054] In this embodiment, since there are non-cooperative targets such as unmanned aerial vehicles and fixed-wing trainers, they may be misjudged as birds; and each cooperative unmanned aerial vehicle and fixed-wing trainer has a GPS device for positioning, so it is less likely to misjudge birds as them; that is, there are mislabeled samples in the bird category, but there are no mislabeled samples in the unmanned aerial vehicle category and the fixed-wing trainer category;

[0055] In this embodiment, the bird samples are input into the classification model trained in step c) to calculate the average loss function value of each bird sample, and the process includes:

[0056] d.1) Each bird sample is divided into standard sub-data sequences, which are respectively used as inputs of the classification model trained in step c);

[0057] d.2) The loss function values of all standard sub-data sequences of the same bird sample ω j are calculated. Wherein, 0<j≤J, J is the total number of bird samples in the training data T in step a), is the number of standard sub-data sequences divided from the bird sample ω j ;

[0058] d.3) The average loss function value of each bird sample is:

[0059] In this embodiment, in step d), the bird samples corresponding to the maximum K average loss function values are mislabeled samples, which are removed from the training data T in step a), and the process includes:

[0060] d.4) arrange E(L(ω j )) from large to small, the loss function mean corresponding to the bird sample in the first K position is the mislabeled sample;

[0061] d.5) remove all the standard sub-data sequences cut out from the mislabeled sample from the training data T in step a);

[0062] In this embodiment, K = [0.1 J], J is the total amount of bird samples in the training data T in step a).

[0063] Preferably, in step e), the real-time identification of the radar target means that the time series data of the target detected by the radar is preprocessed according to step a), is continuously cut into standard sub-data sequences of equal length, and is sequentially sent into the classification model obtained in step e) to calculate the probability of belonging to each category. The probability of the target detected by the radar belonging to a certain category is obtained by averaging the probability of the sub-data sequence cut out from the target in this category. The category with the highest probability is the identification result of the target.

[0064] In this embodiment, γ = 2 in step e), the loss function set not only considers the influence of unbalanced samples, but also gives higher weight to difficult samples, thereby further improving the identification effect.

[0065] Embodiment two;

[0066] Referring to the accompanying Figure 3 A radar target identification device for mislabeled and unbalanced samples, comprising a radar, a computer readable storage medium, a processor, a communication interface and a communication bus; the radar is used for detecting targets, the communication bus communicates between the radar, the computer readable storage medium and the processor through the communication interface, and the computer readable storage medium is used for storing instructions; the processor is used for calling instructions to execute the radar target identification method for mislabeled and unbalanced samples.

[0067] The above embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A radar target recognition method for mislabeling and imbalanced samples, characterized in that, The method comprises the following steps: Step a), pre-processing the labeled target data collected by the radar to obtain all training data T, and dividing all the training data T into a training set Tr and a test set Ts; Step b) building the classification model, training the classification model with the training set Tr of step a) and determining the optimal number of iterations E based on the performance of the classification model on the test set Ts p ; the loss function of the classification model is Where C is the number of sample categories, and p = [p1, ..., p2] C ] is a probability distribution, where each element p i Let y represent the probability that a sample belongs to the i-th class; y = [y1, ..., y2]. C ] is the One-Hot representation of the sample label; when the sample belongs to the i-th class, y i =1, otherwise y i =0; α i It is the weighting coefficient of the i-th class, and its magnitude is inversely proportional to the number of samples in the i-th class; Step c) retraining the classification model built in step b) with all the training data T described in step a), the number of iterations of the training being set to E o ; E o is a positive integer not exceeding the optimal number of iterations E p determined in step b). Step d), inputting a certain class of samples in the all training data T of step a) into the classification model trained in step c), calculating the loss function mean value of each sample of the class, and the samples corresponding to the maximum K loss function mean values are the mislabeled samples in the class, and the mislabeled samples are all removed from the all training data T of step a); wherein K is a predetermined positive integer; step d) is performed for each class of samples to remove all mislabeled samples in all classes of samples; Step e), changing the loss function of the classification model trained in step b) to The training data T is trained again by eliminating the wrongly labeled samples in step d), wherein γ in the loss function is a preset positive number; and the radar target is identified in real time by using the retrained classification model.

2. The radar target recognition method for mislabeling and imbalanced samples according to claim 1, characterized in that, In step a), the target data at least includes high-dimensional time series data composed of target track, speed and radar scattering cross section information; the pre-processing includes interpolation and cleaning of the data, and cutting the high-dimensional time series data into equal-length standard sub-data sequences.

3. The radar target recognition method for mislabeling and imbalanced samples according to claim 1, characterized in that, In step b), the constructed classification model is a deep neural network classification model, and the deep neural network is one of CNN structure, RNN structure, LSTM structure and Transformer structure.

4. The radar target recognition method for mislabeling and imbalanced samples according to claim 3, characterized in that, In the step b), the optimal iteration number E is determined according to the performance of the classification model on the test set Ts p . Specifically, the method is: calculating the auc value of the roc curve of the classification model on the test set Ts at different iteration numbers, and the iteration number corresponding to the maximum auc value is the optimal iteration number E p .

5. The radar target recognition method for mislabeling and imbalanced samples according to claim 1, characterized in that, In step d), the certain class of samples in the all training data T of step a) is input into the classification model trained in step c), and the loss function mean value of each sample of the class is calculated, and the specific method is: d.1), the standard sub-data sequences obtained by cutting each sample in a certain class are input into the classification model trained in step c); d.2) find the loss function value of all standard sub-data sequences of the same sample ω j of the same category where 0 < j ≤ J, J is the total number of samples of the same category in the whole training data T of step a), of the same category j the number of standard sub-data sequences cut out d.3) get the mean of the loss function of each sample in this class as:

6. The radar target recognition method for mislabeling and imbalanced samples according to claim 5, characterized in that, In step d), the samples corresponding to the maximum K loss function mean values are the mislabeled samples in the class, and the mislabeled samples are all removed from the all training data T of step a), and the specific method is: d.4) E(L(ω j )) from large to small, the sample corresponding to the mean of the loss function in the first K positions is the mislabeled sample. d.5) all standard sub-data sequences of mislabeled samples are removed from the training data T of step a).

7. The radar target recognition method for mislabeling and imbalanced samples according to claim 6, characterized in that, In step e), the radar target is identified in real time, and the specific method is: The time series data of the target detected by the radar is pre-processed according to step a), and is continuously cut into equal-length standard sub-data sequences, and is sequentially input into the classification model obtained in step e) to calculate the probability of belonging to each class; the probability of the target detected by the radar belonging to a certain class is obtained by averaging the probability of the sub-data sequence cut from the target in the class; The class with the highest probability is the identification result of the target.

8. The radar target recognition device for mislabeling and imbalanced samples, characterized in that, The radar, the computer readable storage medium, the processor, the communication interface and the communication bus are composed, wherein: The radar is used to detect targets and collect labeled target data; The communication bus communicates between the radar, the computer readable storage medium, the processor through the communication interface; The computer readable storage medium is used to store instructions; The processor is used to call instructions to execute the radar target identification method for mislabeling and unbalanced samples according to any one of claims 1-7.

Images