Intelligent sorting method and system of radar signals based on convolutional denoising autoencoder

Abstract

The present invention discloses a radar signal intelligent sorting method and system based on a convolutional denoising autoencoder, and the method steps are as follows: S1, extracting the pulse TOA sequence of a single radar in an electromagnetic signal database, and sorting the pulse TOA sequence of each radar sample according to the order of their arrival time to construct an aliased pulse sequence; S2, pulse coding the pulse TOA sequence of a single radar and the aliased pulse sequence to obtain a binary coding vector (training data); S3, inputting the obtained binary coding vector into a convolutional denoising autoencoder for training, the output of the model is the binary coding vector of the target pulse, and the trained model is saved; S4, inputting the mixed pulse TOA sequence (real data) received in a complex electromagnetic environment into the trained convolutional denoising autoencoder model to complete signal sorting to obtain the target radar pulse sequence. The present invention improves the accuracy of radar signal sorting and lays a foundation for further processing of radar signal identification, interference, etc.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar reconnaissance, and in particular relates to a radar signal intelligent sorting method and system based on a convolutional denoising autoencoder. The present invention can be used in electronic reconnaissance. Background Art

[0002] Radar signal sorting refers to the process of classifying the radar mixed pulses received by the receiver according to the detected radar pulse parameters and certain prior knowledge according to the real radiation source to which they belong. Only through radar signal sorting can multiple real radar pulse sequences be obtained for further processing such as identification and interference. Radar signal sorting, as the basis of radar reconnaissance signal processing, is an important means of obtaining military intelligence. Research on how to eliminate the negative impact of the electromagnetic environment on radar signals has important national defense and military significance.

[0003] At present, radar signal sorting can be roughly divided into two types. The first is the traditional signal sorting method based on pulse description words (PDW). Pulse description words mainly include pulse width, arrival direction, carrier frequency, pulse amplitude, and pulse arrival time. Among them, pulse repetition interval (PRI) is a parameter with a wide range of variation and more working styles among radar signal parameters, and is the most widely used. It mainly includes cumulative difference histogram method, sequence difference histogram method, PRI transformation method, and plane transformation method. The second is a method based on feature extraction and combined machine learning. Through the transform domain analysis tool, the signal is first transformed in some form, and more feature parameters are mined in the radar pulse signal for sorting, which improves the reliability of radar signal data. The features mainly include some relatively stable intra-pulse features, such as entropy features, phase coefficients, and bispectral features. Finally, various classifier models are used to realize the sorting of radar signals.

[0004] However, the actual battlefield is affected by electromagnetic interference, the signal-to-noise ratio is usually low and there are a lot of TOA estimation errors; on the other hand, the electromagnetic environment is seriously congested, and high-density radar pulses can easily lead to high pulse variance and serious pulse loss at the receiving end, which greatly destroys the laws of radar. At the same time, modern radars tend to be diversified in mode and use, and have flexible working state switching. They can select pulse groups with different parameter combinations and different timing rules according to actual combat conditions, and their early warning and detection capabilities are better than traditional radars. These factors make it difficult for traditional radar signal sorting methods to perform reasonable modeling and statistical analysis based on the laws of pulse characteristics, which brings more severe challenges to signal sorting.

[0005] As a branch or subfield of machine learning, deep learning is one of the latest trends in machine learning and artificial intelligence. Deep learning has been widely used in natural language processing, data mining, computer vision, and autonomous driving due to its powerful feature extraction capabilities. However, there are currently few research results on applying deep learning methods to signal sorting. If the appropriate neural network model and new research ideas can be used to effectively improve the recognition performance of signal sorting in complex electromagnetic environments, it will become another new breakthrough in the field of electronic reconnaissance and have important military significance.

[0006] Based on the above situation, the present invention proposes a radar signal intelligent sorting method and system based on convolutional denoising autoencoder based on deep learning theory. Summary of the invention

[0007] The purpose of the present invention is to improve the radar signal sorting performance in a complex electromagnetic environment, and propose a radar signal intelligent sorting method and system based on a convolutional denoising autoencoder.

[0008] The advantages of the present invention are shown in the following three aspects:

[0009] (1) The present invention uses only one parameter, TOA, which effectively shortens the time for feature extraction. At the same time, in order to make the neural network model handle the increasing pulse arrival time as easily as possible, the TOA sequence is converted into a binary code of only 0 and 1, and the linearly digitized TOA sequence can accept a certain quantization error.

[0010] (2) The convolutional denoising autoencoder used in the present invention is much better than the denoising autoencoder in terms of spatial complexity, greatly reduces the number of training parameters, and effectively prevents overfitting of training.

[0011] (3) Taking into account the changes in parameters such as the missed pulse rate, staggered pulse rate, TOA estimation error, signal-to-noise ratio, and the complex electromagnetic environment in which multifunctional radar signals exist, the sorting accuracy of the present invention is significantly better than that of the PRI transformation method and the SDIF sorting method based on TOA parameters, laying a good foundation for further high-quality processing of radar signal identification and interference.

[0012] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0013] A radar signal intelligent sorting method based on convolutional denoising autoencoder comprises the following steps:

[0014] S1. Extract the pulse TOA sequence of a single radar in the electromagnetic signal database, and sort the pulse TOA sequence of each radar sample according to the order of their arrival time to construct an aliased pulse sequence;

[0015] S2, pulse coding the pulse TOA sequence and aliased pulse sequence of the single radar in step S1 to obtain a binary coding vector (training data);

[0016] S3, input the binary code vector obtained in step S2 into the convolutional denoising autoencoder for training, the output model is the binary code vector of the target pulse, and the trained model is saved;

[0017] S4. Input the mixed pulse TOA sequence (real data) received in a complex electromagnetic environment into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence.

[0018] Further, in step S1, the pulse TOA sequence of a single radar in the electromagnetic signal database is extracted. The TOA sequence of a single radar can be expressed as:

[0019] T={t1,t2,…,t i ,…,t N}

[0020] Among them, t i represents the arrival time of the i-th pulse of a single radar, and N is the number of intercepted pulses.

[0021] Furthermore, in step S1, the pulse TOA sequence of each radar sample is sorted according to the order of their arrival time to construct an aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}, the mixed TOA sequence of multiple radars can be expressed as:

[0022] T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}

[0023] Among them, t ji represents the arrival time of the i-th pulse of the j-th radar, and N represents the number of pulses intercepted by the j-th radar.

[0024] Furthermore, step S2 is specifically: converting the pulse TOA sequence T of the single radar in step S1 into {t1, t2, …, t i ,…,t N} and aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN} respectively perform pulse coding to obtain binary code vectors x and (training data); the pulse encoding process is:

[0025]

[0026] Among them, t unit Represents a given unit, the pulse sequence T according to t unit Linear digitization is performed so that it is in [0,M·t unit ], M is the total number of units in the pulse sequence, i = 0, 1, ... M-1. If a pulse is received within the time window, the vector element corresponding to the time window is represented as 1; if no pulse is received within the time window, the corresponding vector element is represented as 0. M can be expressed as:

[0027] M=t N t unit .

[0028] Further, step S3 is specifically: the obtained binary code vector x and Input the convolutional denoising autoencoder for training. The output of the model is the binary encoding vector z of the target pulse, and the trained model is saved.

[0029] Binary coded vector x and Input to the convolutional denoising autoencoder for model training. The encoding and decoding processes of the convolutional denoising autoencoder are as follows:

[0030]

[0031] z=σ'(W2y+b2)

[0032] Among them, σ and σ' are ReLU activation functions; b1 and b2 are bias vectors; W1 and W2 are weight matrices; x is the original data, is the data polluted by noise; y is the hidden layer feature; z is the output layer data.

[0033] By adjusting the network parameters, the model loss is minimized and the accuracy is high. Since the input x and output z of the model are both binary coded sequences composed of 0 and 1, the binary cross entropy is selected as the objective function:

[0034]

[0035] Furthermore, in step S4, the mixed pulse TOA sequence (real data) received in the complex electromagnetic environment is input into the trained convolutional denoising autoencoder model to complete signal sorting to obtain the target radar pulse sequence.

[0036] The present invention also discloses a radar signal intelligent sorting system based on a convolutional noise reduction autoencoder, comprising the following modules:

[0037] Aliasing pulse sequence construction module: extract the pulse TOA sequence of a single radar in the electromagnetic signal database, and sort the pulse TOA sequence of each radar sample according to the arrival time to construct an aliasing pulse sequence;

[0038] Binary code vector acquisition module: pulse coding the pulse TOA sequence and aliased pulse sequence of a single radar in the aliased pulse sequence construction module to obtain a binary code vector;

[0039] Model output module: The binary coded vector obtained by the binary coded vector acquisition module is input into the convolutional denoising autoencoder for training. The output model is the binary coded vector of the target pulse, and the trained model is saved.

[0040] Signal sorting module: The mixed pulse TOA sequence received in a complex electromagnetic environment is input into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence.

[0041] Preferably, the aliased pulse sequence construction module extracts the pulse TOA sequence T={t1, t2, ..., t i ,…,t N}, the TOA sequence of a single radar is expressed as:

[0042] T={t1,t2,…,t i ,…,t N}

[0043] Among them, t i It represents the arrival time of the i-th pulse of the radar, and N is the number of intercepted pulses.

[0044] Preferably, the aliased pulse sequence construction module arranges the pulse TOA sequence of each radar sample according to the order of their arrival time to construct an aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}, the mixed TOA sequence of multiple radars is expressed as:

[0045] T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}

[0046] Among them, t jirepresents the arrival time of the i-th pulse of the j-th radar, and N represents the number of pulses intercepted by the j-th radar.

[0047] Preferably, in the binary code vector acquisition module, the pulse TOA sequence T of a single radar is converted into {t1, t2, ..., t i ,…,t N} and aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN} respectively perform pulse coding to obtain binary code vectors x and The pulse coding process is:

[0048]

[0049] Among them, t unit Represents a given unit, the pulse sequence T according to t unit Linear digitization is performed so that it is in [0,M·t unit ], M is the total number of units of the pulse sequence, i = 0, 1, ... M-1; if a pulse is received within the time window, the vector element corresponding to the time window is represented by 1; if no pulse is received within the time window, the corresponding vector element is represented by 0; M is represented by:

[0050] M=t N / t unit .

[0051] Preferably, the model output module is as follows: the binary encoding vector x with an input size of 1×784 and As the input of the convolutional denoising autoencoder, the convolutional denoising autoencoder converts the input size into 28×28×1 and performs model training; the encoder part of the convolutional denoising autoencoder consists of two convolutional layers and two pooling layers, and the decoder part consists of two deconvolutional layers and two upsampling layers; the encoding process and decoding process of the convolutional denoising autoencoder are as follows:

[0052]

[0053] z=σ'(W2y+b2)

[0054] Among them, σ and σ' are ReLU activation functions; b1 and b2 are bias vectors; W1 and W2 are weight matrices; x is the original data, is the data polluted by noise; y is the hidden layer feature; z is the output layer data;

[0055] By adjusting the network parameters, the model loss is minimized and the accuracy is high. Since the input x and output z of the model are both binary coded sequences composed of 0 and 1, the binary cross entropy is selected as the objective function:

[0056]

[0057] Compared with the prior art, the present invention comprehensively considers the influence of complex electromagnetic environment, and aims at the technical problem that the sorting performance of the existing radar signal sorting method decreases under the complex electromagnetic environment with multi-function radar signals while considering the changes of parameters such as leakage pulse rate, staggered pulse rate, TOA estimation error, signal-to-noise ratio, etc., and the arrival time of the pulse sequence is encoded and converted into a binary coding vector, and the coding vector is input into a convolutional denoising autoencoder to learn the internal time pattern of the target pulse sequence, and then the mixed pulse sequence is sorted by the trained network to extract the target pulse sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a radar signal sorting method based on a convolutional denoising autoencoder according to Embodiment 1;

[0059] Figure 2 This is a schematic diagram of data processing of an electromagnetic signal database in Example 1;

[0060] Figure 3 This is a schematic diagram of a convolutional denoising autoencoder model in Example 1;

[0061] Figure 4 This is a schematic diagram of signal sorting results under different signal-to-noise ratios in Example 1;

[0062] Figure 5 This is a schematic diagram of signal sorting results under different pulse loss rates in Example 1;

[0063] Figure 6 This is a schematic diagram of signal sorting results under different pulse stagger rates in Example 1;

[0064] Figure 7 This is a schematic diagram of signal sorting results under different TOA estimation errors in Example 1;

[0065] Figure 8 This is a schematic diagram of signal sorting results under different signal-to-noise ratios in an electromagnetic environment with a multifunctional signal in Example 1;

[0066] Fig. 9 It is a schematic diagram of signal sorting results under different TOA estimation errors in an electromagnetic environment with a multifunctional signal in Example 1;

[0067] Fig.10 This is a block diagram of a radar signal sorting system based on a convolutional denoising autoencoder according to Example 2. DETAILED DESCRIPTION

[0068] The following describes the implementation process of the present invention through preferred embodiments, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.

[0069] Embodiment 1

[0070] See also Figure 1 This embodiment provides a radar signal intelligent sorting method based on a convolutional denoising autoencoder, which specifically includes the following steps:

[0071] S1, extracting the pulse TOA sequence of a single radar in the electromagnetic signal database, and then sorting the pulse TOA sequence of each radar sample according to the order of their arrival time to construct an aliased pulse sequence;

[0072] S2, pulse coding the pulse TOA sequence and aliased pulse sequence of the single radar in step S1 to obtain a binary coding vector (training data);

[0073] S3, inputting the binary code vector obtained in step S2 into the convolutional denoising autoencoder for training, the output of the model is the binary code vector of the target pulse, and a trained model is obtained;

[0074] S4. Input the mixed pulse TOA sequence (real data) received in a complex electromagnetic environment into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence.

[0075] Specifically, in step S1, the pulse TOA sequence of a single radar in the electromagnetic signal database is extracted, and then the pulse TOA sequence of each radar sample is sorted according to the order of their arrival time to construct an aliased pulse sequence; in this embodiment, data is extracted from the electromagnetic signal database and preprocessed.

[0076] First, the pulse TOA sequence T = {t1, t2, ..., t i ,…,t N}, the TOA sequence of a single radar can be expressed as:

[0077] T={t1,t2,…,t i ,…,t N}

[0078] Among them, t iIt represents the arrival time of the i-th pulse of the radar, and N is the number of intercepted pulses.

[0079] Then, the pulse TOA sequence of each radar sample is sorted according to the order of their arrival time to construct an aliased pulse sequence T' = {t 11 ,t 12 ,t 31 …,t ji ,…,t jN}, for the specific process, see Figure 2 , the mixed TOA sequence of multiple radars can be expressed as:

[0080] T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}

[0081] Among them, t ji represents the arrival time of the i-th pulse of the j-th radar, and N represents the number of pulses intercepted by the j-th radar.

[0082] In step S2, the pulse TOA sequence and the aliased pulse sequence of the single radar in step S1 are pulse coded to obtain a binary coded vector (training data).

[0083] In this embodiment, data is pulse-coded.

[0084] As the pulse arrival time increases, the TOA sequence eventually increases to a larger value. In order to allow the neural network model to process the input as easily as possible, the TOA sequence of a single radar s(t) is T = {t1, t2, …, t i ,…,t N}Pulse coding is performed to obtain a binary code vector x. The pulse coding process is:

[0085]

[0086] Among them, t unit Represents a given unit, the pulse sequence T according to t unit Linear digitization is performed so that it is in [0,M·t unit ], M is the total number of units in the pulse sequence, i = 0, 1, ... M-1. If a pulse is received within the time window, the vector element corresponding to the time window is represented as 1; if no pulse is received within the time window, the corresponding vector element is represented as 0. M can be expressed as:

[0087] M=t N / t unit .

[0088] In step S3, the obtained binary encoding vector is input into the convolutional denoising autoencoder for training. The output of the model is the binary encoding vector of the target pulse, and the trained model is saved.

[0089] In this embodiment, the model is trained.

[0090] Considering that the convolutional denoising autoencoder is superior to the denoising autoencoder in terms of spatial complexity, it greatly reduces the number of training parameters and effectively prevents overfitting of training. Therefore, the convolutional denoising autoencoder is selected as the model to be trained.

[0091] The binary coded vector x of input size 1×784 and As the input of the convolutional denoising autoencoder, the convolutional denoising autoencoder converts the input size to 28×28×1 and performs model training. The encoder part of the convolutional denoising autoencoder consists of two convolutional layers and two pooling layers, and the decoder part consists of two deconvolutional layers and two upsampling layers. For the specific network structure, see Figure 3 The encoding and decoding process of the convolutional denoising autoencoder is as follows:

[0092]

[0093] z=σ'(W2y+b2)

[0094] Among them, σ and σ' are ReLU activation functions; b1 and b2 are bias vectors; W1 and W2 are weight matrices; x is the original data, is the data polluted by noise; y is the hidden layer feature; z is the output layer data.

[0095] The purpose of model training is to minimize model loss and ensure high accuracy by adjusting network parameters. Therefore, the trained convolutional denoising autoencoder can gradually and accurately sort out the target pulse sequence. They are all binary coded sequences composed of 0 and 1, so binary cross entropy is used as the objective function:

[0096]

[0097] In step S4, the mixed pulse TOA sequence (real data) received in a complex electromagnetic environment is input into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence.

[0098] In this embodiment, pulse sorting is performed.

[0099] The mixed pulse TOA sequence (real data) received in a complex electromagnetic environment is input into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence. The signal sorting results under different electromagnetic environment conditions can be found in Figures 4 to 9 .

[0100] This embodiment comprehensively considers the problem of poor radar signal sorting performance in complex electromagnetic environments. By adding a deep learning algorithm, the accuracy of radar signal sorting can be improved, laying the foundation for further identification, interference and other processing of radar signals.

[0101] Embodiment 2

[0102] like Fig.10 As shown, this embodiment discloses a radar signal intelligent sorting system based on convolutional denoising autoencoder, including the following modules:

[0103] Aliasing pulse sequence construction module: extract the pulse TOA sequence of a single radar in the electromagnetic signal database, and sort the pulse TOA sequence of each radar sample according to the arrival time to construct an aliasing pulse sequence;

[0104] Binary code vector acquisition module: pulse coding the pulse TOA sequence and aliased pulse sequence of a single radar in the aliased pulse sequence construction module to obtain a binary code vector;

[0105] Model output module: The binary coded vector obtained by the binary coded vector acquisition module is input into the convolutional denoising autoencoder for training. The output model is the binary coded vector of the target pulse, and the trained model is saved.

[0106] Signal sorting module: The mixed pulse TOA sequence received in a complex electromagnetic environment is input into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence.

[0107] In this embodiment, the aliased pulse sequence construction module extracts the pulse TOA sequence T = {t1, t2, ..., t i ,…,t N}, the TOA sequence of a single radar is expressed as:

[0108] T={t1,t2,…,t i ,…,t N}

[0109] Among them, t i It represents the arrival time of the i-th pulse of the radar, and N is the number of intercepted pulses.

[0110] In this embodiment, the aliased pulse sequence construction module arranges the pulse TOA sequence of each radar sample according to the order of their arrival time to construct an aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}, the mixed TOA sequence of multiple radars is expressed as:

[0111] T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN}

[0112] Among them, t ji represents the arrival time of the i-th pulse of the j-th radar, and N represents the number of pulses intercepted by the j-th radar.

[0113] In this embodiment, in the binary code vector acquisition module, the pulse TOA sequence T of a single radar is converted into {t1, t2, ..., t i ,…,t N} and aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN} respectively perform pulse coding to obtain binary code vectors x and The pulse coding process is:

[0114]

[0115] Among them, t unit Represents a given unit, the pulse sequence T according to t unit Linear digitization is performed so that it is in [0,M·t unit ], M is the total number of units of the pulse sequence, i = 0, 1, ... M-1; if a pulse is received within the time window, the vector element corresponding to the time window is represented by 1; if no pulse is received within the time window, the corresponding vector element is represented by 0; M is represented by:

[0116] M=t N / t unit .

[0117] In this embodiment, the model output module is as follows: the binary encoding vector x with an input size of 1×784 and As the input of the convolutional denoising autoencoder, the convolutional denoising autoencoder converts the input size into 28×28×1 and performs model training; the encoder part of the convolutional denoising autoencoder consists of two convolutional layers and two pooling layers, and the decoder part consists of two deconvolutional layers and two upsampling layers; the encoding process and decoding process of the convolutional denoising autoencoder are as follows:

[0118]

[0119] z=σ'(W2y+b2)

[0120] Among them, σ and σ' are ReLU activation functions; b1 and b2 are bias vectors; W1 and W2 are weight matrices; x is the original data, is the data polluted by noise; y is the hidden layer feature; z is the output layer data;

[0121] By adjusting the network parameters, the model loss is minimized and the accuracy is high. Since the input x and output z of the model are both binary coded sequences composed of 0 and 1, the binary cross entropy is selected as the objective function:

[0122]

[0123] The present invention aims to improve the accuracy of radar signal sorting by adding a deep learning algorithm to address the problem of poor performance of traditional radar signal sorting methods in complex electromagnetic environments, laying a foundation for further processing of radar signal identification and interference.

[0124] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A radar signal intelligent sorting method based on a convolutional denoising autoencoder, characterized in that: The following steps are involved: S1, extracting the pulse TOA sequence of a single radar in the electromagnetic signal database, and sorting the pulse TOA sequence of each radar sample according to the arrival time to construct an aliased pulse sequence; S2, performing pulse coding processing on the pulse TOA sequence and the aliased pulse sequence of the single radar in step S1 to obtain a binary coding vector; S3, input the binary code vector obtained in step S2 into the convolutional denoising autoencoder for training, the output model is the binary code vector of the target pulse, and the trained model is saved; S4, input the mixed pulse TOA sequence received in the complex electromagnetic environment into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence; Step S1 extracts the pulse TOA sequence T = {t1, t2, ..., t i ,…,t N }, the TOA sequence of a single radar is expressed as: T={t1,t2,…,t i ,…,t N } Among them, t i represents the time when the i-th pulse of the radar arrives, and N is the number of intercepted pulses; Step S1: The pulse TOA sequence of each radar sample is sorted according to the arrival time to construct an aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN }, the mixed TOA sequence of multiple radars is expressed as: T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN } Among them, t ji represents the arrival time of the i-th pulse of the j-th radar, and N represents the number of pulses intercepted by the j-th radar; In step S2, the pulse TOA sequence T of a single radar is converted into {t1, t2, ..., t i ,…,t N } and aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN } respectively perform pulse coding to obtain binary code vectors x and The pulse coding process is: Among them, t unit Represents a given unit, the pulse sequence T according to t unit Linear digitization is performed so that it is in [0,M·t unit ], M is the total number of units of the pulse sequence, i = 0, 1, ... M-1; if a pulse is received within the time window, the vector element corresponding to the time window is represented by 1; if no pulse is received within the time window, the corresponding vector element is represented by 0; M is represented by: M=t N / t unit Step S3, the obtained binary code vector x and Input the convolutional denoising autoencoder for training. The output of the model is the binary encoding vector z of the target pulse. Save the trained model. The details are as follows: The binary coded vector x of input size 1×784 and As the input of the convolutional denoising autoencoder, the convolutional denoising autoencoder converts the input size into 28×28×1 and performs model training; the encoder part of the convolutional denoising autoencoder consists of two convolutional layers and two pooling layers, and the decoder part consists of two deconvolutional layers and two upsampling layers; the encoding process and decoding process of the convolutional denoising autoencoder are as follows: z=σ'(W2y+b2) Among them, σ and σ' are ReLU activation functions; b1 and b2 are bias vectors; W1 and W2 are weight matrices; x is the original data, is the data polluted by noise; y is the hidden layer feature; z is the output layer data; By adjusting the network parameters, the model loss is minimized and the accuracy is high. Since the input x and output z of the model are both binary coded sequences composed of 0 and 1, the binary cross entropy is selected as the objective function:

2. The radar signal intelligent sorting system based on convolutional denoising autoencoder is characterized by: Includes the following modules: Aliasing pulse sequence construction module: extract the pulse TOA sequence of a single radar in the electromagnetic signal database, and sort the pulse TOA sequence of each radar sample according to the arrival time to construct an aliasing pulse sequence; Binary code vector acquisition module: pulse coding the pulse TOA sequence and aliased pulse sequence of a single radar in the aliased pulse sequence construction module to obtain a binary code vector; Model output module: The binary coded vector obtained by the binary coded vector acquisition module is input into the convolutional denoising autoencoder for training. The output model is the binary coded vector of the target pulse, and the trained model is saved. Signal sorting module: input the mixed pulse TOA sequence received in a complex electromagnetic environment into the trained convolutional denoising autoencoder model to complete signal sorting and obtain the target radar pulse sequence; The aliased pulse sequence construction module extracts the pulse TOA sequence T = {t1, t2, ..., t i ,…,t N }, the TOA sequence of a single radar is expressed as: T={t1,t2,…,t i ,…,t N } Among them, t i represents the time when the i-th pulse of the radar arrives, and N is the number of intercepted pulses; The aliased pulse sequence construction module arranges the pulse TOA sequence of each radar sample according to the order of their arrival time to construct an aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN }, the mixed TOA sequence of multiple radars is expressed as: T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN } Among them, t ji represents the arrival time of the i-th pulse of the j-th radar, and N represents the number of pulses intercepted by the j-th radar; In the binary code vector acquisition module, the pulse TOA sequence T of a single radar is converted into {t1, t2, …, t i ,…,t N } and aliased pulse sequence T'={t 11 ,t 12 ,t 31 …,t ji ,…,t jN } respectively perform pulse coding to obtain binary code vectors x and The pulse coding process is: Among them, t unit Represents a given unit, the pulse sequence T according to t unit Linear digitization is performed so that it is in [0,M@t unit ], M is the total number of units of the pulse sequence, i = 0, 1, ... M-1; if a pulse is received within the time window, the vector element corresponding to the time window is represented by 1; if no pulse is received within the time window, the corresponding vector element is represented by 0; M is represented by: M=t N / t unit The model output module is as follows: the input binary encoding vector x of size 1×784 and As the input of the convolutional denoising autoencoder, the convolutional denoising autoencoder converts the input size into 28×28×1 and performs model training; the encoder part of the convolutional denoising autoencoder consists of two convolutional layers and two pooling layers, and the decoder part consists of two deconvolutional layers and two upsampling layers; the encoding process and decoding process of the convolutional denoising autoencoder are as follows: z=σ'(W2y+b2) Among them, σ and σ' are ReLU activation functions; b1 and b2 are bias vectors; W1 and W2 are weight matrices; x is the original data, is the data polluted by noise; y is the hidden layer feature; z is the output layer data; By adjusting the network parameters, the model loss is minimized and the accuracy is high. Since the input x and output z of the model are both binary coded sequences composed of 0 and 1, the binary cross entropy is selected as the objective function:

Images