A radar signal sorting method based on a deep learning BERT model

Abstract

The application provides a radar signal sorting method based on a deep learning BERT model. Parameter information of a pulse description word and corresponding labels are processed to form a data set. The data set is then encoded and divided into a training set, a validation set and a test set. A deep learning BERT model is then built, and the model is trained and verified using the training set and the validation set, and the model parameters are saved. The trained model parameters are then applied to the model for inference on the test set. Finally, the sorting accuracy of the model is calculated, the model parameters with the highest sorting accuracy are retained, and the task model is completed. The application makes full use of more parameters of the PDW, thereby improving the sorting accuracy. The end-to-end mode reduces the participation of manual setting of parameters in sorting. The application is suitable for high-density mixed signal streams and complex modulated signals, and is less affected by complex electromagnetic environments.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of signal and information processing of a radar reconnaissance system, in particular to a radar signal sorting method based on a deep learning BERT model. BACKGROUND

[0002] As one of the key technologies of a radar reconnaissance system, the target of radar signal sorting technology is to separate the pulse sequence generated by each radar from the received random mixed pulse stream. The correctness of the sorting result has a significant impact on the situation awareness ability, information acquisition ability and decision-making ability.

[0003] The pulse description word (PDW) is a radar signal parameter that can be directly obtained by measurement, mainly composed of five instantaneous parameters, namely, radio frequency (RF), pulse width (PW), pulse amplitude (PA), time of arrival (TOA) and direction of arrival (DOA). The pulse repetitive interval (PRI) can be obtained by calculating the TOA, and most of the traditional sorting methods are based on PRI as the starting point for radar signal sorting. The traditional radar signal sorting method based on PRI is actually a deep mining of TOA information, which aims to find radar signals with similar PRI in the mixed pulse stream.

[0004] However, with the diversification of radar modulation types, random modulation and agile modulation have been maturely applied, making it more and more difficult to sort radar signals by relying on a single parameter PRI in the context of increasing pulse density. The use of inter-pulse parameters such as RF, PW, PA and DOA is less. SUMMARY

[0005] The purpose of the present application is to solve the problems of poor adaptability of the traditional radar signal sorting method based on PRI to complex modulation radar signals and poor sorting effect in the context of increasing pulse density in complex electromagnetic environment, and to propose an adaptive signal comprehensive processing technology based on a deep learning BERT model. The technical solution of the present application is:

[0006] Step 1, multiple radars are turned on at the same time, and a receiver is used to obtain mixed radar signals within a period of time to form pulse description words, and the pulse description words are stored in a structured manner.

[0007] Step 2, RF, PW, PA and DOA are normalized and preprocessed;

[0008] Step 3, using known radar information to automatically label the corresponding pulse description word, forming a structured sample set file with column information of RF, PW, PA, TOA, DOA and label;

[0009] Step 4, segmenting the pulse description word in the text data to facilitate the model to learn the parameter characteristics of the pulse description word.

[0010] Step 5, encoding the label in the data set to facilitate the model to update the parameters during training.

[0011] Step 6, dividing the data set, and dividing the training set, the validation set and the test set according to the ratio of 6:2:2.

[0012] Step 7, building a deep learning BERT model, defining the optimizer and the loss function in the model, setting the learning rate and the training rounds, etc.

[0013] Step 8, training the model, saving the model parameters in each training process, and drawing the loss curve on the training set and the validation set to observe whether the model converges.

[0014] Step 9, if the model converges, apply the model parameters saved in all rounds to the test set, calculate the sorting accuracy, draw the accuracy curve, retain the model parameters of the round with the highest sorting accuracy, and complete the task model.

[0015] Compared with the prior art, the present application has the following advantages: (1) compared with radar signal sorting from a single PRI parameter, more parameters of PDW are utilized, thereby improving the sorting accuracy; (2) the end-to-end mode can reduce the participation of manual setting of parameters in sorting; (3) it is suitable for high-density mixed signal stream and complex modulation signal, and is less affected by complex electromagnetic environment. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The sample data processing flowchart provided for the embodiments of the present application;

[0017] Figure 2 The radar signal sorting method based on the deep learning BERT model provided for the embodiments of the present application. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be described in further detail below with reference to the drawings.

[0019] In order to improve the adaptability of the traditional radar signal sorting method based on PRI, the other four parameters in PDW are fully utilized to adapt to the radar signal sorting task requirements in the current complex electromagnetic environment. Deep learning, as a machine learning method based on artificial neural network, has considerable advantages in processing complex problems and data. Deep learning can mine the inherent characteristics of sample data as much as possible, has strong learning ability and generalization ability, and can process high-dimensional data. According to the characteristics of deep learning, the multi-dimensional parameter data of PDW are input into the network model, and the learning ability of the model is used to solve the sorting problem of radar signals.

[0020] Step 1, multiple radars are started at the same time, and a period of time is used to obtain the mixed radar signals to form pulse description words. The RF, PW, PA, TOA and DOA in the pulse description word are stored in a structured manner.

[0021] Step 2, in order to improve the model convergence speed and accuracy to a certain extent, the RF, PW, PA and DOA are normalized and preprocessed.

[0022] Specifically, the preprocessing includes in addition, in order to reflect the possible pulse loss phenomenon in the complex electromagnetic environment, the PDW data of each category in the data set is randomly lost in the same proportion. At the same time, in order to further improve the anti-interference ability of the model, a predetermined number of interference pulses are randomly generated in the sample data, and the same data processing is performed; finally, in order to balance the categories, reduce the model tendency and improve the generalization ability, the number of PDW of each category is kept equal.

[0023] Step 3, write pulse description word batch labeling software, use known radar information to automatically label the corresponding pulse description word, form a structured sample set file with column information of RF, PW, PA, TOA, DOA and label, and the file format is CSV.

[0024] Step 4, load the sample set, which contains pulse description words and label information.

[0025] Step 5, divide the pulse description words in the text data according to the dimensions of RF, PW, PA, TOA and DOA, so as to facilitate the model to learn the parameter characteristics of the pulse description words.

[0026] Step 6, encode the labels in the data set to facilitate parameter updating of the model during training.

[0027] Step 7, divide the data set, divide the training set, validation set and test set according to the preset proportion, and a feasible preset proportion is 6:2:2.

[0028] Step 8, build a deep learning BERT model, define the optimizer and loss function in the model, set the learning rate (learning rate) and training rounds.

[0029] Specifically, the deep learning BERT model adopts a standard BERT model, the number of encoding layers is 12, the optimizer is Adam optimizer, the learning rate is set to 5e-5, the loss function is cross entropy, and the iteration rounds are set to 100. The above parameters are changed according to different actual situations.

[0030] Step 9, train the model, save the model parameters in each training process, add the total loss value of the training set and the total loss value of the validation set calculated in each round to two lists respectively, and draw the loss curve on the training set and the validation set at the end of training to observe whether the model converges.

[0031] Step 10, if the model converges, apply the model parameters saved in all rounds to the test set, calculate the sorting accuracy, and the calculation formula of the accuracy is:

[0032]

[0033] In formula (1), n pp represents the number of predicted results that are category p and the label is also category p, t p represents the real number of category p in the test set; add the accuracy calculated in each round to the list, draw the accuracy curve at the end of inference, keep the model parameters of the round with the highest sorting accuracy, and complete the task model.

[0034] The embodiments of the present application described above do not constitute a limitation on the protection scope of the present application.

Claims

1. A radar signal sorting method based on a deep learning BERT model, characterized in that, The method comprises: Step 1, a plurality of radars are started at the same time, and a pulse description word is formed by simultaneously acquiring mixed radar signals in a period of time using a receiver, and RF, PW, PA, TOA and DOA in the pulse description word are stored in a structured manner; Step 2, RF, PW, PA and DOA are normalized and preprocessed; Step 3, the corresponding pulse description word is automatically labeled using known radar information to form a structured sample set file with column information of RF, PW, PA, TOA, DOA and label; Step 4, load the sample set, wherein the sample set contains pulse description word and label information; Step 5, the pulse description word in the text data is segmented according to the dimensions of RF, PW, PA, TOA and DOA; Step 6, encode the labels in the data set; Step 7, divide the data set, and divide the training set, the validation set and the test set according to the preset proportion; Step 8, build a deep learning BERT model, define the optimizer and the loss function in the model, set the learning rate and the training rounds; Step 9, train the model, save the model parameters in each training process, add the total loss value of the training set and the total loss value of the validation set calculated in each round to two lists respectively, and draw the loss curve on the training set and the validation set at the end of the training to observe whether the model converges; Step 10, if the model converges, apply the saved model parameters in all rounds to the test set to calculate the sorting accuracy.

2. The method of claim 1, wherein, In step 2, the preprocessing includes: Randomly lose the same proportion of PDW data in each category in the data set; Randomly generate a preset number of interference pulses in the sample data and do the same data processing; The number of PDW in each category is kept equal.

3. The method of claim 1, wherein, In step 8, the deep learning BERT model adopts a standard BERT model, the number of encoding layers is 12, the optimizer selects Adam optimizer, the learning rate is set to 5e-5, the loss function is cross entropy, and the iteration rounds are set to 100.

4. The method of claim 1, wherein, The accuracy is determined by the following method: In formula (1), n pp represents the number of prediction results being category p and the label also being category p, t p represents the real number of category p in the test set; the accuracy calculated in each round is added to the list, the accuracy curve is drawn at the end of reasoning, the model parameters of the round with the highest sorting accuracy are retained, and the task model is completed.

Images