Radiation source signal sorting method and device based on feature extraction, equipment and medium

Abstract

The application discloses a radiation source signal sorting method and device based on feature extraction, equipment and a medium, the method comprises the following steps: constructing a pulse arrival time parameter matrix, a pulse repetition interval frequency matrix and a label matrix; the pulse repetition interval frequency matrix is taken as a network input image of a semantic extraction model for training; a semantic feature extraction result predicted by output measured data is output; and signal sorting is completed by extracting a pulse sequence and a staggered sequence. The application has good sorting effect in a mixed scene of staggered PRI modulation and fixed PRI with low signal-to-noise ratio, and is helpful to improve the positioning accuracy of the radiation source subsequently.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radar signal processing, and particularly relates to a radiation source signal sorting method and device based on feature extraction, equipment and a medium. BACKGROUND

[0002] Traditional sorting techniques such as the Sequence Difference Histogram (SDIF) are modified on the basis of the cumulative difference histogram method. The first step needs to calculate the difference histogram of the pulse time of arrival (TOA) of the input pulse stream, and then calculate the detection threshold and obtain the potential PRI. The second step is to search for the corresponding pulse sequence using the potential PRI. SDIF has the problems of poor PRI estimation accuracy and inaccurate sequence extraction for uneven signals in a real complex electromagnetic environment. Although the Pulse Repetition Interval (PRI) transform method has improved accuracy compared with SDIF, the operation complexity is large and is not suitable for dense electromagnetic signal sorting.

[0003] The signal sorting technology based on neural networks mainly identifies the intra-pulse modulation information. In actual scenarios, the amount of intra-pulse sampling point data is large, and the transmission burden is heavy. In engineering, radar signal sorting is usually performed after intra-pulse analysis is completed, and the main parameters used are pulse description words (PDW), including pulse amplitude (PA), pulse frequency (RF), pulse width (PW), pulse time of arrival (TOA) and pulse angle of arrival (DOA). SUMMARY

[0004] The application aims to overcome the defects of the prior art and provide a radiation source signal sorting method, device, equipment and medium based on feature extraction, which aims to mine the characteristics of different PRI modulation types from the pulse sequence time parameters, i.e. TOA and PRI, to realize fast and accurate extraction of target signal characteristics and high-precision sorting.

[0005] The application achieves the above-mentioned purpose through the following technical solutions.

[0006] A radiation source signal sorting method based on feature extraction, the method comprising:

[0007] Constructing a pulse time of arrival parameter matrix, a pulse repetition interval frequency matrix and a label matrix;

[0008] Training the pulse repetition interval frequency matrix as a network input image of a semantic extraction model;

[0009] Output measured data prediction semantic feature extraction result;

[0010] Extracting pulse sequence and staggered sequence completes signal sorting.

[0011] Further, the construction of the pulse arrival time parameter matrix specifically includes:

[0012] Calculate the time difference between any two pulse arrival times in the radar signal pulse arrival time sequence and take the positive value to obtain the pulse arrival time difference matrix.

[0013] Further, the construction of the pulse repetition interval frequency matrix specifically includes:

[0014] Generate the pulse repetition interval frequency matrix based on the pulse arrival time difference matrix.

[0015] Further, the construction of the label matrix specifically includes:

[0016] Construct a label matrix consistent in size with the pulse repetition interval frequency matrix, and divide the signal pulse repetition interval modulation type with numbers at each position.

[0017] Further, the semantic extraction model specifically includes a full convolutional network based on an encoding-decoding structure, the full convolutional network is connected through skip connection, and the skip path of the full convolutional network is as follows:

[0018]

[0019] Where H(h) is the convolution operation and the activation function operation, U(·) represents the up-sampling layer, [·] represents the connection layer, x i,j is the jth node of the ith layer, the node with j=0 is the first layer input of the encoding network, the input of the node with j=1 is the output of the consecutive two layers of the encoding network, and the node with j>1 has j+1 inputs, wherein the jth node input is the output of the first j nodes in the same skip path, and the up-sampling output from the lower level;

[0020] The full convolutional network calculates the loss function of the outputs of all branches and takes the average, and the full convolutional network sets the loss function for the outputs of four semantic levels, which is a combination of binary cross entropy and dice coefficient:

[0021]

[0022] Where L is the loss function, Y, represents the prediction probability and the true label of a batch, Y b , represents the prediction probability and the true label of the bth image, and N represents the batch size.

[0023] Further, the output semantic feature extraction result of the measured data prediction specifically comprises:

[0024] The measured staggered pulse sequence is truncated to n to obtain a pulse arrival time sequence {t1, t2,..., tn}; n};

[0025] According to D TOA (i,j)=t j -t i The pulse arrival time difference matrix corresponding to the measured staggered pulse sequence is obtained The matrix size is n*n;

[0026] The frequency of the elements in the pulse arrival time difference matrix is counted to obtain a corresponding pulse repetition interval frequency matrix;

[0027] The pulse repetition interval frequency matrix is input into a semantic extraction model to obtain a predicted semantic feature extraction result matrix

[0028] Further, the signal sorting by extracting the pulse sequence and the uneven sequence specifically comprises:

[0029] The pulse arrival time sequence corresponding to the stable pulse repetition interval and the pulse arrival time sequence corresponding to the uneven pulse repetition interval are extracted;

[0030] The pulse arrival time sequence corresponding to the stable pulse repetition interval comprises recording the position index (i,j) with a label of 1 in the semantic feature extraction result matrix , considering that the pulse repetition interval corresponding to (i,j) is the interval between t i and t j , extracting all positions of the same pulse repetition interval, and sorting by time to obtain the pulse arrival time sequence of the received single radiation source emitted pulse, thereby completing signal sorting;

[0031] The pulse arrival time sequence corresponding to the uneven pulse repetition interval specifically comprises:

[0032] The position index (i,j) with a label of 2 in the semantic feature extraction result is recorded, and the pulse arrival time sequence extracted by a certain pulse repetition interval value is {t1, t4, t6, t 10 , t 11 , t 15 ...}, and the corresponding time slice sequence is {{t1, t4}, {t6, t 10}, {t 11 , t 15},...} By using the merging principle based on whether the time slices overlap or not, the pulse repetition interval value combination from the same radiation source is established, and in a certain staggered pulse repetition interval combination, if the time slices corresponding to all pulse repetition intervals in the frame do not overlap with each other and the proportion in the observation time of the whole pulse arrival time is higher than a threshold, it is considered that the corresponding staggered pulse repetition interval comes from the same radiation source, and vice versa.

[0033] In another aspect, the present application also provides a radiation source signal sorting device based on feature extraction, the device comprising:

[0034] A matrix establishing module is configured to construct a pulse arrival time parameter matrix, a pulse repetition interval frequency matrix and a label matrix;

[0035] A model training module is configured to train the pulse repetition interval frequency matrix as a network input image;

[0036] A feature extraction module is configured to output a semantic feature extraction result predicted by the measured data;

[0037] A signal sorting module is configured to extract a pulse sequence and a staggered sequence to complete signal sorting.

[0038] In another aspect, the present application also provides a computer device comprising a processor and a memory, wherein the memory stores a computer program, the computer program is loaded and executed by the processor to realize any one of the above-mentioned radiation source signal sorting methods based on feature extraction.

[0039] In another aspect, the present application also provides a computer readable storage medium, wherein the storage medium stores a computer program, the computer program is loaded and executed by a processor to realize any one of the above-mentioned radiation source signal sorting methods based on feature extraction.

[0040] The present application has the following advantages:

[0041] (1) The present application realizes high-precision sorting of radar radiation source signals, and compared with the traditional sequence difference histogram algorithm and PRI transformation method, it can estimate accurate PRI values and overcome the cumulative error caused by sequence search.

[0042] (2) Compared with the existing sorting method based on deep network, the present application only uses the time parameter TOA, which can not only overcome the problem of poor PW or RF measurement accuracy, but also can mine the dynamic change characteristics of the sequence in the time dimension. This method has good sorting effect in the mixed scene of staggered PRI modulation and fixed PRI with low signal-to-noise ratio, which helps to improve the positioning accuracy of the subsequent radiation source. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 This is a schematic diagram of the radiation source signal sorting method based on feature extraction provided in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the semantic extraction model architecture according to an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of time slice merging analysis according to an embodiment of the present invention;

[0046] Figure 4 This is a structural block diagram of a radiation source signal sorting device based on feature extraction provided in an embodiment of the present invention. Detailed Implementation

[0047] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0048] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Traditional methods, when applied to complex electromagnetic environments, suffer from poor PRI estimation accuracy and inaccurate sequence extraction for staggered signals. Neural network-based signal sorting techniques primarily identify intra-pulse modulation information; however, in real-world scenarios, the amount of data from intra-pulse sampling points is large, resulting in a heavy transmission burden.

[0050] To address the aforementioned technical problems, the following embodiments of the radiation source signal sorting method, apparatus, device, and medium based on feature extraction of the present invention are proposed.

[0051] Example 1

[0052] Reference Figure 1 ,like Figure 1 The diagram shows a flowchart of the radiation source signal sorting method based on feature extraction provided in this embodiment. The method specifically includes the following steps:

[0053] Step 1: Construct the TOA difference matrix D TOA Let the radar signal TOA sequence be {t1, t2, ..., t...} n},n is the total number of pulses in the sequence, the time difference between any two TOAs is calculated and only the positive value is taken, an n x n TOA difference matrix D TOA Each element in the TOA difference matrix D

[0054] D TOA (i,j) = t j -t i ;

[0055] The elements in the TOA difference matrix D

[0056]

[0057] Step 2: Construct the PRI frequency matrix F PRI The PRI frequency matrix is generated on the basis of the TOA difference matrix D TOA , the matrix size is n x n, and the element F PRI (i,j) in the matrix corresponds to the frequency of the value of D TOA (i,j). The PRI frequency matrix obtained in this step is the image to be processed in the present application, which is also the input image of the network.

[0058] Step 3: Construct the label matrix L PRI The image semantic feature extraction task is to classify each pixel. Therefore, the size of the label matrix should be consistent with the size of the generated PRI frequency matrix, and the PRI modulation type is divided by a number at each position. For example, set 0 for the background, set 1 for the fixed PRI position, set 2 for the staggered signal, set 3 for the PRI that does not belong to the above two types and is not 0, and filter out all the second harmonics and pulse group repetition intervals (PGRI). The second harmonic is defined as twice the real PRI, and when the degree of pulse loss is large, the PRI may be estimated as its multiple harmonics. The PGRI is the sum of the sub-periods of the staggered PRI signal, which is expressed as:

[0059]

[0060] where PRI i is the i-th PRI value in a certain PRI combination of a staggered radiation source, and m is the number of sub-PRI of the PRI combination of the radiation source. The label matrix of the training set is constructed in this step, and the PRI category division method is consistent with the PRI category form after the actual data prediction.

[0061] Step 4: Input the PRI frequency matrix F PRIThe network is trained as input image. This step aims to obtain a semantic feature extraction model for the PRI frequency matrix, so that the measured data can obtain more accurate category prediction results. The network structure used is a fully convolutional network based on an encoding-decoding structure, and the encoding and decoding networks are connected through a series of redesigned nested and dense Skip-connections, aiming to reduce the semantic gap between the decoding results and the feature maps before encoding. The skip connection path is as follows:

[0062]

[0063] where H(·) is the convolution operation and the activation function operation, U(·) represents the up-sampling layer, [·] represents the connection layer, x i ,j is the i-th layer of the j-th node. The node of j=0 is the first layer input of the encoding network, the input of the node of j=1 is the output of the consecutive two layers of the encoding network, and the node of j>1 has j+1 inputs, wherein the output of the j nodes is the output of the first j nodes in the same skip connection path, and the up-sampling output from the lower level.

[0064] Due to the nested skip connection path, the network generates full-resolution feature maps {x 0,j , j e {1, 2, 3, 4}} at multiple semantic levels, and the network calculates the loss function for the outputs of all branches and takes the average. The model sets the loss function for the outputs of the four semantic levels as a combination of binary cross-entropy and dice coefficient:

[0065]

[0066] where L is the loss function, Y, represents the prediction probability and the true label of a batch, Y b , represents the prediction probability and the true label of the b-th image, and N represents the batch size.

[0067] Step 5: Output the semantic feature extraction result of the measured data prediction. This step obtains the semantic feature extraction result of the PRI frequency matrix of the measured data based on the semantic feature extraction model obtained after training. The length of the measured staggered pulse sequence is truncated to n, and the TOA sequence is obtained as {t1, t2,..., t n}. According to the D TOA (i, j) = t j -t i The sequence corresponding to the pulse arrival time difference matrix is obtained. The size of the matrix is n x n. The frequency of the elements in is counted to obtain the corresponding PRI frequency matrix The matrix size is also n x n. The PRI frequency matrix is input into the trained model in step 4 to obtain a predicted semantic feature extraction result matrix

[0068] Step 6: Extract the pulse sequence and staggered sequence analysis.

[0069] (1) Extract the TOA sequence corresponding to the stable PRI

[0070] The semantic feature extraction result matrix The position index (i, j) with a label of 1 is recorded, and it is considered that the PRI corresponding to the position is t i and the interval between t j , all positions with the same PRI are extracted, and the TOA sequence of the received single radiation source emitted pulse is obtained by sorting by time, and the sorting task is realized.

[0071] (2) Extract the TOA sequence corresponding to the staggered PRI

[0072] The position index (i, j) with a label of 2 in the semantic feature extraction result is recorded, and it is assumed that the TOA sequence extracted by a certain PRI value is {t1, t4, t6, t 10 , t 11 , t 15 ...}, and the time slice sequence corresponding to the PRI is {{t1, t4}, {t6, t 10}, {t 11 , t 15},...}. Further, the combination of PRI values from the same radiation source is established by using the merging principle based on whether the time slices overlap or not. Within a certain staggered PRI combination, if the time slices corresponding to all PRIs in the frame do not overlap with each other and have a high proportion in the entire TOA observation time, it is considered that the staggered PRI combination comes from the same radiation source, and otherwise it is considered that the combination is not established.

[0073] Compared with the traditional sequence difference histogram algorithm and PRI transformation method, the embodiment can estimate accurate PRI values and overcome the cumulative error caused by sequence search. Compared with the existing sorting method based on deep network, the present scheme only uses the time parameter TOA, which can not only overcome the problem of poor measurement accuracy of PW or RF, but also can mine the dynamic change characteristics of the sequence in the time dimension.

[0074] Embodiment 2

[0075] In this embodiment, 1100 samples containing noise and mixed staggered and fixed PRIs are sequentially subjected to data preprocessing.

[0076] Step 1: Cut the pulse sequence with a length of 1024 to construct the TOA difference matrix D TOALet the radar signal TOA sequence be {t1, t2, ..., t...} 1024 Calculating the time difference between two adjacent TOAs and only taking positive values ​​yields a 1024×1024 TOA difference matrix, which is an upper triangular matrix. The PRI range is set to 100μs-2000μs, with elements outside this range set to 0. The TOA difference matrix D... TOA Each element in the table is:

[0077] D TOA (i,j)=t j -t i ;

[0078] The elements in this matrix satisfy the following relationship:

[0079]

[0080] Step 2: Construct the PRI frequency matrix F PRI The PRI frequency matrix is ​​derived from the difference matrix D of TOA. TOA The matrix is ​​generated based on the above, with a size of 1024×1024, and the elements F in the matrix are... PRI (i,j) corresponds to D TOA The frequency of (i,j) values. The PRI frequency matrix obtained in this step is the image to be used for semantic feature extraction in this invention, and also the input image of the network.

[0081] Step 3: Construct the label matrix L PRI The image semantic feature extraction task involves classifying each pixel. Therefore, the size of the label matrix should be consistent with the size of the generated PRI frequency matrix, which is 1024×1024, and the signal PRI modulation type should be numerically assigned at each position. For example, the background is set to 0, fixed PRI positions are set to 1, staggered signals are set to 2, and PRIs that do not belong to either category and are non-zero are set to 3. All second harmonics and pulse group repetition intervals (PGRI) are filtered out. The second harmonic is defined as twice the true PRI; when the pulse loss is significant, the PRI may be estimated as its multiple harmonics. PGRI is the sum of the sub-periods of the staggered PRI signal, expressed as:

[0082]

[0083] Among them PRI i Let be the i-th PRI value in a certain staggered PRI combination, and m be the number of sub-PRIs in this PRI combination. This step constructs the label matrix of the training set, and the PRI category classification method is consistent with the PRI category form after prediction from the actual data.

[0084] Step 4: The PRI frequency matrix F PRI is trained with network input images. The training set is 1000 images and the test set is 100 images. The 10-fold cross-validation method is used. The network structure used is a fully convolutional network based on an encoding-decoding structure, which is described in Figure 2 Figure 2 The semantic extraction model architecture diagram of this embodiment is shown in the figure. The encoding and decoding networks are connected through a series of redesigned nested and dense Skip-connections, aiming to reduce the semantic gap between the decoding results and the feature maps before encoding. The skip connection path is as follows:

[0085]

[0086] where H(·) is the convolution operation and the activation function operation, U(·) represents the up-sampling layer, [·] represents the connection layer, x i ,j is the i-th layer j-th node. The node with j = 0 is the input of the first layer of the encoding network, the input of the node with j = 1 is the output of the two consecutive layers of the encoding network, and the node with j > 1 has j+1 inputs, in which the j-th node input is the output of the first j nodes in the same skip connection path, and the up-sampling output from the lower level.

[0087] Due to the nested skip connection paths, the network generates full-resolution feature maps {x 0,j , j ∈ {1, 2, 3, 4}} at multiple semantic levels. The network calculates the loss function for the outputs of all branches and takes the average. The model sets the loss function for the outputs of the four semantic levels as a combination of binary cross-entropy and dice coefficient:

[0088]

[0089] where L is the loss function, Y, represents the predicted probability and the true label of a batch, Y b , represents the predicted probability and the true label of the b-th image, and N represents the batch size.

[0090] Experiments show that part of the background in the semantic feature extraction result of a sample is divided into noise, but it does not affect the extraction of effective PRI. By comparing the actual label matrix, it can be concluded that the network prediction has obtained a more accurate classification result.

[0091] ​Step 5: Output the semantic feature extraction result of the measured data prediction. This step is based on the semantic feature extraction model obtained after training, and the PRI frequency matrix obtained by processing the measured data. The measured staggered pulse sequence is truncated to a length of 1024 to obtain the TOA sequence as {t1, t2,..., t 1024}. According to the D TOA (i, j) = t j -t i , the sequence corresponding to the pulse arrival time difference matrix is obtained The frequency of the elements in is counted to obtain the corresponding PRI frequency matrix The PRI frequency matrix is input into the model trained in step 4 to obtain the predicted semantic feature extraction result matrix The dimensions of the matrix and are both 1024 x 1024.

[0092] Step 6: Extract the pulse sequence and analyze the irregular sequence

[0093] The positions with labels 1 and 2 in the semantic feature extraction result are extracted and counted, and the final statistical result is obtained after removing the invalid PRI with low frequency and the higher harmonics of the remaining valid PRI

[0094] (1) Extract the TOA sequence corresponding to the stable PRI

[0095] The position index (i, j) with label 1 in the semantic feature extraction result matrix is recorded, and the PRI corresponding to this position is considered to be the interval between t i and t j . All positions with the same PRI are extracted and sorted by time to obtain the TOA sequence of the received single radiation source emitted pulse, and the sorting task is realized.

[0096] (2) Extract the TOA sequence corresponding to the irregular PRI

[0097] The position index (i, j) with label 2 in the semantic feature extraction result is recorded, and the TOA sequence extracted for a certain PRI value is set as {t1, t4, t6, t 10 , t 11 , t 15 ...}. The time slice sequence corresponding to this PRI is {{t1, t4}, {t6, t 10}, {t 11 , t 15},...}. Further, the merging principle based on whether the time slices overlap is used to determine the PRI value combination from the same radiation source. Referring toFigure 3 As shown in Figure 3 The time slice merging analysis schematic diagram of this embodiment is shown in the figure. In a certain PRI combination, if all the time slices corresponding to the PRIs in the frame do not overlap with each other and account for a high proportion in the entire TOA observation time, it is considered that the PRI combination comes from the same radiation source, and vice versa, it is considered that the combination is not established.

[0098] Embodiment 3

[0099] Referring to Figure 4 As shown in Figure 4 The structure block diagram of the radiation source signal sorting device based on feature extraction provided by this embodiment is shown in the figure, and the device specifically includes the following structures:

[0100] The matrix establishing module is used for constructing a pulse arrival time parameter matrix, a pulse repetition interval frequency matrix and a label matrix;

[0101] The model training module trains the pulse repetition interval frequency matrix as a network input image;

[0102] The feature extraction module outputs the semantic feature extraction result of the measured data prediction;

[0103] The signal sorting module extracts the pulse sequence and the PRI sequence to complete the signal sorting.

[0104] Embodiment 4

[0105] The preferred embodiment provides a computer device, which can realize the steps in any embodiment of the radiation source signal sorting method based on feature extraction provided by the embodiments of the present application, and thus can realize the beneficial effects of the radiation source signal sorting method based on feature extraction provided by the embodiments of the present application. Details are described in the previous embodiments, which will not be repeated here.

[0106] Embodiment 5

[0107] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by related hardware controlled by the instructions. The instructions can be stored in a computer readable storage medium and loaded and executed by a processor. Therefore, the embodiments of the present application provide a storage medium, which stores a plurality of instructions. The instructions can be loaded by a processor to execute the steps in any embodiment of the radiation source signal sorting method based on feature extraction provided by the embodiments of the present application.

[0108] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, or the like.

[0109] Due to the instructions stored in the storage medium, the steps in any feature extraction-based radiation source signal sorting method embodiment provided by the embodiments of the present application can be executed, thus achieving the beneficial effects of any feature extraction-based radiation source signal sorting method provided by the embodiments of the present application. Details are described in the foregoing embodiments, which will not be repeated here.

[0110] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of signal sorting of a radiation source based on feature extraction, characterized in that, The method comprises: constructing a pulse arrival time parameter matrix, a pulse repetition interval frequency matrix, and a label matrix; training the pulse repetition interval frequency matrix as a network input image of a semantic extraction model; outputting a semantic feature extraction result predicted by the measured data; extracting pulse sequences and uneven sequences to complete signal sorting; constructing the pulse arrival time parameter matrix specifically comprises: calculating the time difference between any two pulse arrival times in the radar signal pulse arrival time sequence and taking the positive value to obtain a pulse arrival time difference matrix; constructing the pulse repetition interval frequency matrix specifically comprises: generating the pulse repetition interval frequency matrix based on the pulse arrival time difference matrix; The semantic extraction model specifically comprises a fully convolutional network based on an encoding-decoding structure, the fully convolutional network is connected through skip connections, and the skip path of the fully convolutional network is as follows: ； wherein is a convolution operation and an activation function operation, denotes an up-sampling layer, denotes a concatenation layer, is the first layer of the encoding network, is the first layer of the encoding network, is the first layer of the encoding network, is the first layer of the encoding network, is the first layer of the encoding network, has inputs, wherein the node input is the output of the previous nodes in the same hop path, and up-sampling output from lower levels; The fully convolutional network calculates the loss function of the outputs of all branches and takes the average, and the fully convolutional network sets the loss function for the outputs of the four semantic levels as a combination of binary cross entropy and dice coefficients: ； wherein, is a loss function, denotes predicted probabilities and true labels for a batch, denotes predicted probabilities and true labels for the th image, denotes the batch size.

2. The method of claim 1, wherein, Constructing the label matrix specifically comprises: constructing a label matrix consistent in size with the pulse repetition interval frequency matrix, and dividing the signal pulse repetition interval modulation type by numbers at each position.

3. The feature extraction based radiation source signal sorting method of claim 1, wherein, The semantic extraction model specifically comprises a fully convolutional network based on an encoding-decoding structure, the fully convolutional network is connected through skip connections, and the skip path of the fully convolutional network is as follows: ； in It involves convolution operations and activation function operations. Indicates the upsampling layer. Indicates the connection layer. For the first Layer 1 node The node is the first layer input of the encoding network. The input to a node is the output of two consecutive layers of the encoding network. have There are 1 input, of which The node input is the previous one in the same jump path The output of each node, as well as the upsampled output from the lower level; The fully convolutional network calculates the loss function of the outputs of all branches and takes the average, and the fully convolutional network sets the loss function for the outputs of the four semantic levels as a combination of binary cross entropy and dice coefficients: ； wherein, is a loss function, denotes predicted probabilities and true labels for a batch, denotes predicted probabilities and true labels for the th image, denotes the batch size.

4. The method of claim 1, wherein, The output semantic feature extraction result predicted by the measured data specifically comprises: The measured staggered pulse sequence length is truncated as , and the pulse arrival time sequence is ; According to obtaining a pulse arrival time difference matrix corresponding to the measured staggered pulse sequence , the matrix size is ; Counting the frequency of occurrence of elements in the matrix of pulse arrival time differences to obtain a corresponding pulse repetition interval frequency matrix; The pulse repetition interval frequency matrix is input into a semantic extraction model to obtain a predicted semantic feature extraction result matrix .

5. The method of claim 4, wherein, The extraction of pulse sequences and uneven sequences to complete signal sorting specifically comprises: extracting the pulse arrival time sequence corresponding to the stable pulse repetition interval and the pulse arrival time sequence corresponding to the uneven pulse repetition interval; Wherein, the extracting the pulse arrival time sequence corresponding to the stable pulse repetition interval includes extracting the semantic feature extraction result matrix The position index with the label 1 Recorded, consider The corresponding pulse repetition interval is And The interval between them, all the same pulse repetition interval positions are extracted, and sorted by time to obtain the pulse arrival time sequence of the received single radiation source emission pulse, completing signal sorting; extracting the pulse arrival time sequence corresponding to the uneven pulse repetition interval specifically comprises: The position index with the label 2 in the semantic feature extraction result Recorded, suppose the pulse repetition interval value extracted from the pulse arrival time sequence is Then the corresponding time slice sequence is Wherein m≤n, using the merging principle based on whether the time slices overlap or not, the pulse repetition interval value combination from the same radiation source is established, in a certain staggered pulse repetition interval combination, if the time slices corresponding to all pulse repetition intervals in the frame do not overlap with each other, and the proportion in the entire pulse arrival time observation time is higher than the threshold, it is considered that the corresponding staggered pulse repetition interval comes from the same radiation source, otherwise it is considered to come from different sources.

6. A feature extraction based emitter signal sorting apparatus applying the feature extraction based emitter signal sorting method as claimed in claim 1, characterized by, The device comprises: a matrix establishing module, which is configured to construct a pulse arrival time parameter matrix, a pulse repetition interval frequency matrix, and a label matrix; a model training module, which trains the pulse repetition interval frequency matrix as a network input image; a feature extraction module, which outputs a semantic feature extraction result predicted by the measured data; a signal sorting module, which extracts pulse sequences and uneven sequences to complete signal sorting.

7. A computer device, characterized by The computer device comprises a processor and a memory, the memory stores a computer program, the computer program is loaded and executed by the processor to implement the feature extraction-based radiation source signal sorting method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, the computer program is loaded and executed by the processor to implement the feature extraction-based radiation source signal sorting method according to any one of claims 1 to 5.

Images