阅读说明：本技术 基于改进K-Means联合卷积自编码器的信号分选方法及装置 (Signal sorting method and device based on improved K-Means combined convolution self-encoder ) 是由 李鹏 申慧芳 武斌 张葵 田卫东 张东燕 郭瑞鹏 于 2021-07-19 设计创作，主要内容包括：本发明公开了一种基于改进K-Means联合卷积自编码器的信号分选方法,包括：获取不同PRI调制类型的雷达对应的若干TOA序列；构建卷积自编码器,并利用TOA序列对其进行训练,得到若干训练好的卷积自编码器；对不同PRI调制类型的雷达信号进行仿真,得到混合PDW序列；基于改进的K-Means聚类算法对混合PDW序列进行预分选,得到预分选后的TOA序列；利用训练好的卷积自编码器对预分选后的TOA序列进行脉冲提取,得到信号分选结果。本发明提供的信号分选方法适用于复杂环境下的雷达信号分选,提高了信号分选准确率和分选效率。(The invention discloses a signal sorting method based on an improved K-Means combined convolution self-encoder, which comprises the following steps: acquiring a plurality of TOA sequences corresponding to radars with different PRI modulation types; constructing a convolution self-encoder, and training the convolution self-encoder by utilizing a TOA sequence to obtain a plurality of trained convolution self-encoders; simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence; pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence; and performing pulse extraction on the pre-sorted TOA sequence by using a trained convolution self-encoder to obtain a signal sorting result. The signal sorting method provided by the invention is suitable for radar signal sorting in a complex environment, and improves the signal sorting accuracy and sorting efficiency.)

1. A signal sorting method based on an improved K-Means joint convolution self-encoder, comprising:

acquiring a plurality of TOA sequences corresponding to radars with different PRI modulation types;

constructing a convolution self-encoder, and training the convolution self-encoder by using the TOA sequence to obtain a plurality of trained convolution self-encoders;

simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence;

pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence;

and performing pulse extraction on the pre-sorted TOA sequence by utilizing a trained convolution self-encoder to obtain a signal sorting result.

2. The method of claim 1, wherein the obtaining the TOA sequences corresponding to different types of PRI modulation radar comprises:

respectively simulating radar TOA data of modulation types of a fixed PRI, a group-variation PRI, a spread PRI, a jitter PRI and a sliding-variation PRI to obtain corresponding TOA sequences; wherein the TOA sequence has a proportion of interference pulses.

3. The signal sorting method based on the improved K-Means joint convolution self-encoder according to claim 1, wherein the convolution self-encoder is constructed and trained by using the TOA sequence to obtain a plurality of trained convolution self-encoders, and the method comprises:

respectively constructing an encoder network and a decoder network;

connecting the encoder network and the decoder network into a convolutional self-encoder network, and setting corresponding network parameters;

encoding the TOA sequence;

and respectively inputting the coded TOA sequences serving as training sets into the convolutional auto-encoder network, and iteratively updating the network weight for a certain number of times to obtain a plurality of trained convolutional auto-encoders.

4. The method of claim 3, wherein the encoder network comprises 3 layers of one-dimensional convolution, and the structure of the convolution is: the device comprises a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a third convolution layer, a third pooling layer, a flat layer, a first full-connection layer and a coding output layer;

the decoder network comprises 4 layers of one-dimensional convolution, and the structure of the decoder network is as follows in sequence: the decoding device comprises a second full connection layer, a third full connection layer, a first up-sampling layer, a fourth convolution layer, a second up-sampling layer, a fifth convolution layer, a third up-sampling layer, a sixth convolution layer, a fourth up-sampling layer and a decoding output layer.

5. The signal sorting method based on the improved K-Means joint convolution auto-encoder according to claim 3, wherein the TOA sequence is encoded by:

setting a certain interval unit;

and when the TOA sequence is judged to be within a certain interval unit, setting the interval value to be 1, otherwise, setting the interval value to be 0.

6. The signal sorting method based on the improved K-Means joint convolution self-encoder as claimed in claim 1, wherein after obtaining the mixed PDW sequence, the mixed PDW sequence is preprocessed before pre-sorting the mixed PDW sequence to map the pulse parameters to the same parameter variation unit.

7. The signal sorting method based on the improved K-Means joint convolution self-encoder according to claim 1, wherein the pre-sorting of the mixed PDW sequence based on the improved K-Means clustering algorithm to obtain the pre-sorted TOA sequence comprises:

taking a certain pulse sequence of the mixed PDW sequence as an initial clustering center;

calculating the distance between the current pulse sequence and the clustering center according to the PW and the RF of the current sequence, clustering the current pulse sequence according to the distance, and updating the clustering center at the same time;

circularly clustering each pulse sequence with all clustering centers until all pulse sequences are traversed;

and counting the number of samples of each cluster, and deleting isolated points to obtain a pre-sorted TOA sequence.

8. The signal sorting method based on the improved K-Means joint convolution self-encoder according to claim 7, wherein the step of calculating the distance between the current pulse sequence and the cluster center according to the PW and RF of the current sequence, clustering the current pulse sequence according to the distance, and updating the cluster center comprises the steps of:

if a certain cluster center is cen ═ a, b, then the current cluster center isThe distance between the pulse sequence i and the cluster center is:

wherein, PW_iIndicating the pulse width, RF, of the current ith pulse_iThe carrier frequency of the current ith pulse is represented;

if the d is smaller than the preset distance threshold value, distributing the current pulse to the clustering center, searching samples in the class at the same time, and averaging all the samples to update the clustering center; otherwise, adding the current ith pulse as a new cluster center.

9. The signal sorting method based on the improved K-Means joint convolution self-encoder according to claim 1, wherein the trained convolution self-encoder is used to perform pulse extraction on the pre-sorted TOA sequence to obtain a signal sorting result, and the method comprises:

binary coding the pre-sorted TOA sequence;

and sequentially inputting the coded data into the trained convolution autoencoders, and constructing a pulse retrieval formula to perform pulse extraction to obtain a signal sorting result.

10. A signal sorting apparatus based on an improved K-Means joint convolution auto-encoder, comprising:

the system comprises a first data acquisition module (1) for acquiring a plurality of TOA sequences corresponding to radars of different PRI modulation types;

the model building module (2) is used for building a convolution self-encoder and training the convolution self-encoder by utilizing the TOA sequence to obtain a plurality of trained convolution self-encoders;

the second data acquisition module (3) is used for simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence;

the pre-sorting module (4) is used for pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence;

and the pulse extraction module (5) is used for performing pulse extraction on the pre-sorted TOA sequence by utilizing a trained convolution self-encoder to obtain a signal sorting result.

Technical Field

The invention belongs to the technical field of signal processing, and particularly relates to a signal sorting method and device based on an improved K-Means combined convolution self-encoder.

Background

Electronic warfare, also called electronic countermeasure, is an important combat means of modern warfare, and is mainly various electronic measures and actions taken by two opposing parties to weaken and destroy the use efficiency of electronic equipment of the other party and ensure the performance of the electronic equipment of the own party. For a long time, radar countermeasure has been dominant in electronic warfare, and radar reconnaissance is one of the main contents of radar countermeasure. In general, the signals intercepted by a radar reconnaissance system include, in addition to a large number of radar signals, other electromagnetic radiation source signals that are randomly mixed with a plurality of radar signals to form an interleaved signal stream. During subsequent processing, each radar signal needs to be separated from a signal stream according to data such as intercepted radar characteristic parameters, arrival time and arrival direction, then parameter extraction and radar model identification are carried out on the selected signals, and finally information such as radar type, attribute, application and threat degree is obtained according to identification results. The radar radiation source signal sorting is a key link of radar reconnaissance signal processing, directly influences the performance of radar reconnaissance equipment, and is related to subsequent combat decisions. Therefore, the technical level of signal sorting has become an important mark for measuring the advanced technology of the radar reconnaissance system.

At present, the existing radar signal sorting methods mainly have two categories, one is the traditional radar signal sorting method, and the signal sorting is mainly realized by taking PRI (pulse repetition interval) as a main sorting characteristic parameter. The other type is a multi-parameter radar signal sorting method which mainly utilizes various intra-pulse parameters as sorting characteristic parameters to realize signal sorting.

However, the conventional radar signal sorting method is only suitable for radar signals with a conventional inter-pulse modulation mode, and cannot meet the radar signal sorting requirement in a complex environment, while the conventional multi-parameter radar signal sorting method has an unsatisfactory sorting effect on pulse sequences with obvious lost pulses and more clutters, and has low sorting efficiency, and wrong or untimely signal sorting can cause a large amount of batch increase and batch omission (corresponding to false alarms and missed alarms), thereby seriously affecting the countermeasure effect and even the success or failure of the battle. Therefore, the radar signal sorting method which is suitable for complex environments and has a good sorting effect is designed to have a crucial significance for radar countermeasure.

Disclosure of Invention

In order to solve the above problems in the prior art, the present invention provides a signal sorting method and apparatus based on an improved K-Means joint convolution auto-encoder. The technical problem to be solved by the invention is realized by the following technical scheme:

a signal sorting method based on an improved K-Means joint convolution self-encoder, comprising:

acquiring a plurality of TOA sequences corresponding to radars with different PRI modulation types;

constructing a convolution self-encoder, and training the convolution self-encoder by using the TOA sequence to obtain a plurality of trained convolution self-encoders;

simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence;

pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence;

and performing pulse extraction on the pre-sorted TOA sequence by utilizing a trained convolution self-encoder to obtain a signal sorting result.

In an embodiment of the present invention, the obtaining a plurality of TOA sequences corresponding to radars of different PRI modulation types includes:

respectively simulating radar TOA data of modulation types of a fixed PRI, a group-variation PRI, a spread PRI, a jitter PRI and a sliding-variation PRI to obtain corresponding TOA sequences; wherein the TOA sequence has a proportion of interference pulses.

In an embodiment of the present invention, constructing a convolutional auto-encoder, and training the convolutional auto-encoder by using the TOA sequence to obtain a plurality of trained convolutional auto-encoders, including:

respectively constructing an encoder network and a decoder network;

connecting the encoder network and the decoder network into a convolutional self-encoder network, and setting corresponding network parameters;

encoding the TOA sequence;

and respectively inputting the coded TOA sequences serving as training sets into the convolutional auto-encoder network, and iteratively updating the network weight for a certain number of times to obtain a plurality of trained convolutional auto-encoders.

In one embodiment of the present invention, the encoder network comprises 3 layers of one-dimensional convolution, whose structure is in turn: the device comprises a first convolution layer, a first pooling layer, a second convolution layer, a second pooling layer, a third convolution layer, a third pooling layer, a flat layer, a first full-connection layer and a coding output layer;

the decoder network comprises 4 layers of one-dimensional convolution, and the structure of the decoder network is as follows in sequence: the decoding device comprises a second full connection layer, a third full connection layer, a first up-sampling layer, a fourth convolution layer, a second up-sampling layer, a fifth convolution layer, a third up-sampling layer, a sixth convolution layer, a fourth up-sampling layer and a decoding output layer.

In one embodiment of the present invention, the coding of the TOA sequence includes:

setting a certain interval unit;

and when the TOA sequence is judged to be within a certain interval unit, setting the interval value to be 1, otherwise, setting the interval value to be 0.

In an embodiment of the present invention, after obtaining the mixed PDW sequence, before pre-sorting the mixed PDW sequence, the mixed PDW sequence needs to be preprocessed to map the pulse parameters to the same parameter variation unit.

In an embodiment of the present invention, the pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence, includes:

taking a certain pulse sequence of the mixed PDW sequence as an initial clustering center;

calculating the distance between the current pulse sequence and the clustering center according to the PW and the RF of the current sequence, clustering the current pulse sequence according to the distance, and updating the clustering center at the same time;

circularly clustering each pulse sequence with all clustering centers until all pulse sequences are traversed;

and counting the number of samples of each cluster, and deleting isolated points to obtain a pre-sorted TOA sequence.

In one embodiment of the present invention, calculating a distance between the current pulse sequence and a cluster center according to the PW and RF of the current sequence, clustering the current pulse sequence according to the distance, and updating the cluster center at the same time, includes:

and if a certain cluster center is cen ═ a, b, the distance between the current pulse sequence i and the cluster center is:

wherein, PW_iIndicating the pulse width, RF, of the current ith pulse_iThe carrier frequency of the current ith pulse is represented;

if the d is smaller than the preset distance threshold value, distributing the current pulse to the clustering center, searching samples in the class at the same time, and averaging all the samples to update the clustering center; otherwise, adding the current ith pulse as a new cluster center.

In an embodiment of the present invention, performing pulse extraction on the pre-sorted TOA sequence by using a trained convolutional auto-encoder to obtain a signal sorting result, including:

binary coding the pre-sorted TOA sequence;

and sequentially inputting the coded data into the trained convolution autoencoders, and constructing a pulse retrieval formula to perform pulse extraction to obtain a signal sorting result.

Another embodiment of the present invention provides a signal sorting apparatus based on an improved K-Means joint convolution self-encoder, including:

the system comprises a first data acquisition module 1, a second data acquisition module and a third data acquisition module, wherein the first data acquisition module is used for acquiring a plurality of TOA sequences corresponding to radars with different PRI modulation types;

the model establishing module 2 is used for establishing a convolution self-encoder and training the convolution self-encoder by utilizing the TOA sequence to obtain a plurality of trained convolution self-encoders;

the second data acquisition module 3 is used for simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence;

the pre-sorting module 4 is used for pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence;

and the pulse extraction module 5 is used for performing pulse extraction on the pre-sorted TOA sequence by using a trained convolution self-encoder to obtain a signal sorting result.

The invention has the beneficial effects that:

1. the method has the advantages that the convolution self-encoder is introduced to learn the internal mode of each radar pulse sequence, the output sequence is generated from the self-encoder with good training, the good sorting effect is realized on the pulse sequences with obvious lost pulses and more clutters, the method is suitable for sorting radar signals in a complex environment, and the signal sorting accuracy is improved;

2. according to the invention, the K-Means clustering algorithm is improved, so that the clustering K value can be automatically determined according to the set distance threshold and the updating of the mean clustering center, and the sorting efficiency is improved while the radar signal sorting requirement is met.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

FIG. 1 is a schematic flow chart of a signal sorting method based on an improved K-Means joint convolution self-encoder according to an embodiment of the present invention;

fig. 2 is a schematic structural diagram of a signal sorting device based on an improved K-Means joint convolution self-encoder according to an embodiment of the present invention.

Detailed Description

The present invention will be described in further detail with reference to specific examples, but the embodiments of the present invention are not limited thereto.

Example one

Referring to fig. 1, fig. 1 is a schematic flow chart of a signal sorting method based on an improved K-Means joint convolution auto-encoder according to an embodiment of the present invention, which includes:

step 1: and acquiring a plurality of TOA sequences corresponding to the radars with different PRI modulation types.

Generally, Pulse Description Word (PDW) is often used for multi-parameter radar signal sorting, and mainly includes parameters such as Pulse Width (PW), Direction Of Arrival (DOA), Carrier Frequency (CF), Pulse Amplitude (PA), and Time Of Arrival (TOA), which can well describe the internal characteristics Of pulses and become important sorting characteristic parameters in a proper order. Among them, TOA is one of the important parameters for accurately locating the position of the transmitter.

Different TOA sequences may result in different self-encoder performance indicators. Therefore, the present embodiment selects several TOA sequences of different PRI modulation types as the input of the whole algorithm to obtain different self-encoder network models.

Typical PRI modulation types mainly include five types, namely a fixed PRI, a group-varying PRI, a staggered PRI, a jittering PRI and a sliding PRI, and the five modulation types basically cover most radar signals, so that a plurality of TOA sequences corresponding to the radars of the five PRI modulation types are selected as input data in the present embodiment, which indicates that the method provided by the present embodiment can adapt to different types of radar signals.

Specifically, in this embodiment, the MATLAB environment is used to simulate the TOA data of the above five PRI modulation types, so as to obtain five corresponding TOA sequences, and each modulation signal is added with an interference pulse in a certain proportion. The simulation parameters of the five types of PRI modulation are respectively as follows:

the repetition period of the fixed PRI modulation radar is 92us, the number of pulses is 800, the pulse loss rate is 10 percent, the measurement error is 0.1us, and 50 percent of interference pulse is added;

the repetition period of the group-change PRI modulated radar is 510us, 534us, 545us and 578us, the number of the group-change PRI modulated radar is 6, the number of the pulses is 144, the pulse loss rate is 5 percent, the measurement error is 0.1us, and 50 percent of interference pulses are added;

the repetition period of the staggered PRI modulated radar is 310us, 417us, 533us, 577us, 587us, 687us, 723us and 823us, the number of pulses is 128, the pulse loss rate is 5 percent, the measurement error is 0.1us, and 50 percent of interference pulses are added;

the repetition period of the sliding PRI modulation radar is 190us-208us, 3us is taken as increment, the number of pulses is 371, the pulse loss rate is 10 percent, the measurement error is 0.1us, and 50 percent of interference pulse is added;

the central repetition period of the jitter PRI modulation radar is 380us, the jitter rate is 10%, the number of pulses is 190, the pulse loss rate is 10%, the measurement error is 0.1us, and 50% of interference pulses are added;

and (3) simulating by using an MATLAB simulation tool according to the parameter setting to obtain five TOA sequences, wherein each sequence comprises 10000 groups of pulse data.

Step 2: constructing a convolution self-encoder, and training the convolution self-encoder by utilizing a TOA sequence to obtain a plurality of trained convolution self-encoders; the method specifically comprises the following steps:

21) an encoder network and a decoder network are constructed separately.

Firstly, a coder network containing 3 layers of one-dimensional convolution is built, and the structure of the coder network is as follows in sequence: the first convolution layer → the first pooling layer → the second convolution layer → the second pooling layer → the third convolution layer → the third pooling layer → the flat layer → the first fully-connected layer → the encoded output layer.

Further, the number of convolution kernels in the first to third convolution layers is respectively set to 16, 32 and 64, the sizes of the convolution kernels are respectively set to 31 × 1, 31 × 1 and 16 × 1, the step sizes are all set to 2, the activation function is a relu function, the first to third pooling layers all adopt a maximum pooling mode, the sizes of the pooling region kernels are respectively set to 2 × 1, 2 × 1 and 2 × 1, the step sizes are all set to 2, the node of the first full-link layer is 64, the activation function is a relu function, the node of the coding output layer is 32, and the activation function is a relu function.

Specifically, the mathematical model of the activation function as relu function is represented as follows:

where f (x) represents the response of the input value x of the network after the activation function relu.

Then, a decoder network containing 4 layers of one-dimensional convolution is built, and the structure of the decoder network sequentially comprises the following steps: second full-link layer → third full-link layer → first upsampled layer → fourth convolutional layer → second upsampled layer → fifth convolutional layer → third upsampled layer → sixth convolutional layer → fourth upsampled layer → decoded output layer.

Further, the second fully-connected layer node is set to 64, the activation function is a relu function, the third fully-connected layer node is set to 3008, and the activation function is a relu function; the upsampling factors of the first to fourth upsampling layers are sequentially set to be 2, 4, 4 and 2; the number of convolution kernels in the fourth convolution layer, the sixth convolution layer and the fourth convolution layer is respectively set to be 64, 32 and 16, the sizes of the convolution kernels are respectively set to be 8 multiplied by 1, 8 multiplied by 1 and 31 multiplied by 1, the step sizes are all set to be 1, and the activation function is a relu function; the decoding output layer is a one-dimensional convolution layer, the number of convolution kernels is set to be 1, the size of the convolution kernels is set to be 9 x 1, and the activation function is a relu function.

22) And connecting the encoder network and the decoder network into a convolutional self-encoder network, and setting corresponding network parameters.

In this embodiment, the loss function of the formed convolutional autoencoder network can be set to binary _ cross, the optimizer to adam, the number of training batches to 60, and the learning rate to 0.001.

23) The TOA sequence is encoded.

In this embodiment, before the convolutional autocoder is trained by using the TOA sequence, the TOA sequence needs to be binary coded.

Specifically, a small interval unit t may be set_spaceWhen the value of the TOA sequence is judged to be in a certain interval unit, the value of the interval is set to be 1, otherwise, the value of the interval is set to be 0, and the formula expression is as follows:

wherein, t_iRepresenting the value of the ith sequence.

The five TOA sequences obtained by the coding method are respectively coded to respectively obtain corresponding five binary code streams.

24) And respectively inputting the coded TOA sequences serving as training sets into a convolutional self-encoder network, and iteratively updating the network weight for a certain number of times to obtain a plurality of trained convolutional self-encoders.

In the embodiment, the Adam algorithm is adopted to iteratively update the network weight, and the convolutional auto-encoder is considered to be trained when the number of updating times reaches 100 times.

Specifically, the five binary code streams obtained in the third step are used as training samples to be respectively input into the trained convolutional self-encoders, and the Adam algorithm is used for iteratively updating the weights of the network for 100 times, so that the five trained convolutional self-encoder networks are obtained.

And step 3: and simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence.

In order to simulate the actually received radar signals as much as possible, the present embodiment simulates the mixed signals of multiple radars in the MATLAB environment, and obtains a mixed PDW sequence. Some radars have similar parameters in carrier frequency and pulse width dimensions.

Specifically, based on the five modulation types selected in step 1, the five radars are still selected here for simulation, and the pulse descriptor parameters of each radar are as follows:

the radar I: pulse width is 6us, carrier frequency is 1800MHz, and TOA sequence is the sequence parameter of the fixed radar in the step 1;

and a second radar: the pulse width is 6.5us, the carrier frequency is 1820MHz, and the TOA sequence is the sequence parameter of the group-change radar in the step 1;

radar III: the pulse width is 9us, the carrier frequency is 1913MHz, and the TOA sequence is the sequence parameter of the difference radar in the step 1;

and fourthly, radar: the pulse width is 10us, the carrier frequency is 1660MHz, and the TOA sequence is the sequence parameter of the slip radar in the step 1;

and V, radar: the pulse width is 12.4us, the carrier frequency is 1533MHz, and the TOA sequence is the sequence parameter of the jittered radar in the step 1.

And rearranging the PDW sequences of the five radars according to the arrival time to obtain a mixed PDW sequence.

It should be noted that, the parameters of the description words of the radar signal pulses received by the receiver are very complex and the measurement units are different. Wherein, the variation range of the pulse width is 0.2-50 us, and the variation range of the carrier frequency is 2-18 GHz. In order to determine the similarity relationship between pulses, pulse parameters need to be preprocessed to map the pulse parameters to the same parameter change unit, and the cluster analysis is performed with the same magnitude.

Therefore, after obtaining the mixed PDW sequence, the mixed PDW sequence needs to be preprocessed before being pre-sorted.

Firstly, a matrix of a PDW sequence needs to be normalized, each dimension parameter is mapped to a variation range with a mean value of 0 and a standard deviation of 1, so as to achieve the effect of dimension elimination, and the formula is as follows:

wherein p is_ikIs the k-dimension pulse parameter of the ith pulse, n is the number of pulses,and s_kRespectively mean value and standard deviation p of each dimension parameter in radar signal pulse description word matrix_i'_kTo representNormalized pulse parameters.

Further, in order to map the mapping range into the [0, 1] range, the extremization process is performed again to obtain a normalized extremization matrix. The polarization process is shown as follows:

wherein, max (p)_i'_k)、min(p_i'_k) Respectively representing the maximum and minimum standard values, p, of each dimensional parameter in the standardized matrix_i”_kIndicating the pulse parameters subjected to the polarization processing.

In addition, if the parameter variation of the input signal is almost the same, the dimensional parameter fluctuates within the same value, the standard deviation approaches to 0, and the maximum value and the minimum value of the dimensional parameter obtained through standardization are almost equal, so that the extremization processing cannot be performed. At this time, a certain tolerance range can be set for each dimension parameter, when the variation range of the dimension parameter is smaller than the set tolerance range, the dimension parameter is considered to have small variation and small difference, and the extreme value processing process can be skipped at this time.

And 4, step 4: and pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence.

41) One pulse sequence of the mixed PDW sequence is used as an initial cluster center.

Preferably, this embodiment may use the first pulse, i.e. the first sample, of the mixed PDW sequence after the normalized polarization preprocessing as the initial clustering center.

42) And calculating the distance between the current pulse sequence and the clustering center according to the PW and the RF of the current sequence, clustering the current pulse sequence according to the distance, and updating the clustering center at the same time.

Specifically, if a certain cluster center is cen ═ (a, b), the distance between the current pulse sequence i and the cluster center is:

wherein, PW_iIndicating the pulse width, RF, of the current ith pulse_iRepresenting the carrier frequency of the current i-th pulse.

If the d is smaller than the preset distance threshold thre, distributing the current pulse to the clustering center, searching samples in the class at the same time, and averaging all the samples to update the clustering center; otherwise, adding the current ith pulse as a new cluster center.

43) And circularly clustering each pulse sequence with all the clustering centers until all the pulse sequences are traversed.

44) And counting the number of samples of each cluster, and deleting isolated points to obtain a pre-sorted TOA sequence.

And 5: and performing pulse extraction on the pre-sorted TOA sequence by utilizing a trained convolution self-encoder to obtain a signal sorting result.

51) And carrying out binary coding on the pre-sorted TOA sequence.

Specifically, the pre-sorted TOA sequence is binary-coded here with reference to the coding scheme in step 23).

52) And sequentially inputting the coded data into a plurality of trained convolution self-encoders, and constructing a pulse retrieval formula to perform pulse extraction to obtain a signal sorting result.

Specifically, the encoded data is sent to a trained network, and the value of the output vector is a decimal between 0 and 1 after the sigmoid function. The target output x has a value of 0 or 1, where 0 indicates that a pulse was not correctly retrieved and 1 indicates that a pulse was correctly retrieved. This embodiment can determine the accuracy and measure the accurate loss of the network by setting a threshold τ, which is expressed as follows:

where τ is always set to 0.5 to ensure a balanced distribution of the spurious and target pulses.

Finally, an output matrix x is obtained, which is the same as the input dimension, where the position of 1 in the matrix is the pulse for which the PRI is correctly retrieved. Assuming that the element at the ith position in x is 1, the pulse search formula is as follows:

wherein, t_spaceFor the interval unit set in step 23), the range of the corresponding TOA sequence is represented as (t)_is,t_ie) The pulses in the time interval are the pulses for which the PRI was correctly retrieved.

And (3) respectively utilizing the five convolutional encoders obtained in the step (2) to perform pulse selection on the encoded data so as to obtain a matching result.

This embodiment learns every radar pulse sequence's internal mode through introducing the convolution self-encoding ware, generates the output sequence from training good self-encoding ware, has good sorting effect to losing the pulse obvious and the more pulse sequence of clutter, is applicable to the radar signal under the complex environment and selects separately, has improved the signal and has selected separately the rate of accuracy.

Meanwhile, the K-Means clustering algorithm is improved, so that the clustering K value can be automatically determined according to the set distance threshold value and the updating of the mean clustering center, and the sorting efficiency is improved while the radar signal sorting requirement is met.

Example two

On the basis of the first embodiment, the embodiment provides a signal sorting device based on the improved K-Means joint convolution self-encoder. Referring to fig. 2, fig. 2 is a schematic structural diagram of a signal sorting apparatus based on an improved K-Means joint convolution auto-encoder according to an embodiment of the present invention, which includes:

the system comprises a first data acquisition module 1, a second data acquisition module and a third data acquisition module, wherein the first data acquisition module is used for acquiring a plurality of TOA sequences corresponding to radars with different PRI modulation types;

the model building module 2 is used for building a convolution self-encoder and training the convolution self-encoder by utilizing the TOA sequence to obtain a plurality of trained convolution self-encoders;

the second data acquisition module 3 is used for simulating radar signals of different PRI modulation types to obtain a mixed PDW sequence;

the pre-sorting module 4 is used for pre-sorting the mixed PDW sequence based on an improved K-Means clustering algorithm to obtain a pre-sorted TOA sequence;

and the pulse extraction module 5 is used for performing pulse extraction on the pre-sorted TOA sequence by using the trained convolutional self-encoder to obtain a signal sorting result.

The apparatus provided in this embodiment can implement the signal sorting method provided in the first embodiment, and the detailed process is not described herein again.

The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.