Radiation source target association method, device, medium and system based on depth auto-encoder

Abstract

The invention discloses a radiation source target correlation method, device, medium and system based on a depth auto-encoder, and belongs to the field of electromagnetic signal processing, and the method comprises the steps: extracting the sequence of five dimensions, namely frequency, repetition frequency, pulse width, amplitude and orientation, of a clustering result of each processing beat, and obtaining a clustering result of each processing beat; then, constructing a deep auto-encoder based on a CNN in combination with an LSTM network to perform feature representation of a clustering result; in the process of realizing radiation source target association by the deep auto-encoder reasoning application, firstly, a trained deep auto-encoder network is adopted for clustering results of all processing beats to obtain feature representation vectors; and then whether two corresponding clustering results belong to the same radiation source target is judged based on the joint loss of different processing beat feature representation vectors, so that stable correlation of the radiation source target is realized. According to the invention, the robustness and adaptability of complex agile waveform radiation source target sensing can be improved.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic signal processing, and more specifically, to a method, apparatus, medium, and system for correlating radiation sources and targets based on a depth autoencoder. Background Technology

[0002] As radiation source targets increasingly adopt complex and agile waveform signals, traditional radiation source target association methods based on electromagnetic parameter matching become increasingly inadequate, especially under complex electromagnetic environments and dynamic observation conditions. For example, traditional target association algorithms rely on electromagnetic parameters such as repetition frequency and carrier frequency. When radiation source targets adopt agile waveforms, their repetition frequencies and carrier frequencies become completely non-repeating, causing traditional methods based on electromagnetic parameter matching to fail. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a radiation source target association method, device, medium and system based on a deep autoencoder, which can improve the robustness and adaptability of complex and agile waveform radiation source target perception.

[0004] The objective of this invention is achieved through the following approach: A radiation source-target association method based on a deep autoencoder includes the following steps: For the clustering results of each processing beat, the sequences of five dimensions, namely frequency, repetition frequency, pulse width, amplitude, and orientation, are extracted. Then, a deep autoencoder based on CNN combined with LSTM network is used to represent the features of the clustering results. The CNN network represents the local features of the clustering results, and the LSTM network represents the temporal features of the clustering results. In the process of realizing radiation source target association in the deep autoencoder inference application, the clustering results of each processing beat are first obtained by using a trained deep autoencoder network to obtain feature representation vectors. Then, based on the joint loss of the feature representation vectors of different processing beats, it is determined whether the two corresponding clustering results belong to the same radiation source target, thereby realizing stable association of radiation source targets.

[0005] Furthermore, the training loss function of the deep autoencoder network is a joint loss function constructed by combining mean squared error and cosine similarity loss.

[0006] Furthermore, the feature representation of the clustering results based on the deep autoencoder constructed using a combination of CNN and LSTM networks, where the CNN network represents the local features of the clustering results and the LSTM network represents the temporal features of the clustering results, specifically includes the following sub-steps: Let the clustering result of the nth clustering in the mth processing beat be represented as: ,in This represents the frequency of the k-th pulse in the clustering result. This represents the repetition frequency of the k-th pulse in the clustering result. This represents the pulse width of the k-th pulse in the clustering result. This represents the amplitude of the k-th pulse in the clustering result. This indicates the location of the k-th pulse in the clustering result. M represents the maximum number of processing cycles, N represents the maximum number of clusters in each processing cycle, and K represents the maximum number of pulses in each cluster result; CNN networks are composed of stacked convolutional blocks consisting of convolution operations, activation function processing, batch normalization, and pooling. The specific operation steps are as follows: ; in, This indicates convolution processing. This indicates activation function processing. This indicates normalization processing. Indicates pooling processing; , This indicates the number of times convolutional blocks are stacked in a CNN network; After constructing the CNN network, an LSTM network is used to further extract the temporal features of the clustering network. The formula is as follows: ; in, This represents the hidden layer of the LSTM network at the last moment of time series processing, serving as the feature representation of the corresponding deep autoencoder input; Then, with As input, the constructed deep autoencoder is used for processing to obtain the feature representation vectors of the clustering results in each processing beat. ,set up ,in express Dimensions These represent the parameters for each dimension.

[0007] Furthermore, the training of the deep autoencoder network specifically includes the following sub-steps: The constructed deep autoencoder network is trained by randomly selecting two sample pairs from the training set for processing, and constructing a joint loss based on mean squared error and cosine similarity loss. At the same time, the Adam optimizer is used to train and obtain the parameters of the deep autoencoder network.

[0008] Furthermore, the clustering results for each processing beat are used to obtain feature representation vectors using a trained deep autoencoder network. Then, based on the joint loss of the feature representation vectors of different processing beats, it is determined whether the two corresponding clustering results belong to the same radiation source target. This specifically includes the following sub-steps: For clustering results with different processing rhythms, the clustering results are first processed by the trained deep autoencoder to obtain the feature representation of the corresponding clustering results. Then, for the feature representation vectors of different processing rhythms, the joint loss based on mean square error and cosine similarity is calculated respectively, and a threshold is set to determine whether the two clustering results belong to the same radiation source target.

[0009] Further, the step of calculating the joint loss based on mean square error and cosine similarity, and setting a threshold to determine whether the two clustering results belong to the same radiation source target, specifically includes the following sub-steps: Let the feature representation vector of the nth clustering result in the mth processing beat be... The feature representation vector of the nth clustering result in the (m+1)th processing beat is The threshold is set to Thres, and the joint loss is... The sub-steps for determining whether two clustering results belong to the same radiation source target are as follows: 1) If only one is found satisfy If the two clustering results represent the same radiation source target, then the nth cluster of the (m+1)th processing beat is directly associated with the nth cluster of the mth processing beat as a radiation source target; 2) If multiple results are found satisfy , then it means It may be correlated with multiple clustering results; in this case, choose... Associate the smallest cluster results; 3) If all None of them satisfy less than or equal to If the result is not the same as any of the clustering results, then a new radiation source target should be created and reported.

[0010] A radiation source target association device based on a depth autoencoder includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method described in any of the preceding methods.

[0011] A computer-readable storage medium storing a computer program that, when loaded by a processor, executes the method described in any of the preceding claims.

[0012] A radiation source target association system based on a depth autoencoder includes the radiation source target association device based on a depth autoencoder as described above.

[0013] The beneficial effects of this invention include: This invention proposes a radiation source target association method based on a deep autoencoder. For the clustering results of each processing beat, sequences of five dimensions—frequency, repetition rate, pulse width, amplitude, and azimuth—are extracted. Then, a deep autoencoder based on a CNN combined with an LSTM network is used to represent the feature representation of the clustering results. The CNN network represents the local features of the clustering results, and the LSTM network represents the temporal features. The loss function used for network training is a joint loss function constructed from mean squared error and cosine similarity loss. In the process of radiation source target association through deep autoencoder inference, the trained deep autoencoder network is first used to obtain feature representation vectors for the clustering results of each processing beat. Then, based on the joint loss of the feature representation vectors of different processing beats, it is determined whether two corresponding clustering results belong to the same radiation source target, thereby achieving stable association of radiation source targets. This invention's method constructs a deep autoencoder by combining local and temporal features, and then improves the robustness and adaptability of complex and agile waveform radiation source target perception processing through deep feature representation of clustering results and large-scale training. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the method according to an embodiment of the present invention. Detailed Implementation

[0016] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0017] In view of the problems in the background, in a preferred embodiment, a radiation source-target association method based on a deep autoencoder is particularly provided, such as Figure 1 As shown, it includes the following steps: For the clustering results of each processing beat, sequences with five dimensions—frequency, repetition rate, pulse width, amplitude, and orientation—are extracted. Then, a deep autoencoder based on a CNN combined with an LSTM network is used to represent the features of the clustering results. The CNN network represents the local features of the clustering results, while the LSTM network represents the temporal features. The loss function used for network training is a joint loss function constructed from mean squared error and cosine similarity loss. In the process of achieving radiation source target association through deep autoencoder inference, the trained deep autoencoder network is first used to obtain feature representation vectors for the clustering results of each processing beat. Then, based on the joint loss of the feature representation vectors of different processing beats, it is determined whether two corresponding clustering results belong to the same radiation source target, thereby achieving stable association of radiation source targets. This embodiment's method constructs a deep autoencoder by combining local and temporal features, and then improves the robustness and adaptability of complex and agile waveform radiation source target perception processing through deep feature representation of clustering results and large-scale training.

[0018] In other embodiments, based on the methods of the above embodiments, a radiation source target association method based on a depth autoencoder is further provided, which includes the following steps: Suppose that the clustering result of the nth clustering in the mth processing cycle is represented as follows: ,in This represents the frequency of the k-th pulse in the clustering result. This represents the repetition frequency of the k-th pulse in the clustering result. This represents the pulse width of the k-th pulse in the clustering result. This represents the amplitude of the k-th pulse in the clustering result. This indicates the location of the k-th pulse in the clustering result. M represents the maximum number of processing cycles, N represents the maximum number of clusters in each processing cycle, and K represents the maximum number of pulses in each clustering result. The method of this embodiment of the invention includes the following steps: Step 1: Construct a deep autoencoder based on a CNN combined with an LSTM network to represent the clustering results. The CNN network represents the local features of the clustering results, while the LSTM network represents the temporal features. The CNN network consists of stacked convolutional blocks formed by convolution operations, activation function processing, batch normalization, and pooling. The specific steps are as follows:

[0019] in This indicates convolution processing. This indicates activation function processing. This indicates normalization processing. Indicates pooling processing; , This indicates the number of times convolutional blocks are stacked in a CNN network.

[0020] Following the CNN network, an LSTM network is used to further extract the temporal features of the clustering network, as shown in the formula:

[0021] in This represents the hidden layer of the LSTM network at the last moment of the time series processing, serving as the feature representation of the deep autoencoder for that input.

[0022] In this embodiment, the CNN network used here is an improved ResNet14 network. The input is 6*K, and the output is 32*K / 64; the LSTM network output is 32*1, and the 32*1 feature vector here is the feature representation vector of the clustering result.

[0023] Step Two: With As input, the deep autoencoder constructed in step one is used for processing to obtain the feature representation vectors of the clustering results in each processing beat. Assuming ,in express Dimensions These represent the parameters for each dimension.

[0024] Step 3: Train the constructed deep autoencoder network by randomly selecting two sample pairs from the training set for processing, and constructing a joint loss based on mean squared error and cosine similarity loss; at the same time, use the Adam optimizer to train and obtain the parameters of the deep autoencoder network.

[0025] Suppose that the feature representation vectors obtained by two randomly selected sample pairs through a deep autoencoder are as follows: and .

[0026] Its mean square error loss is: , The cosine similarity loss is:

[0027] in Indicates the first One dimension, Indicates the first The expected value of each dimension feature.

[0028] The combined loss is: , here , indicating the equal weighting of the two types of losses.

[0029] Step 4: In the process of applying deep autoencoder inference to radiation source target association, for the clustering results of different processing beats, the clustering results are first processed using the trained deep autoencoder to obtain the feature representation of the clustering results. Then, for the feature representation vectors of different processing beats, their joint loss based on mean square error and cosine similarity is calculated respectively, and a threshold is set to determine whether two clustering results belong to the same radiation source target.

[0030] Suppose the feature representation vector of the nth clustering result in the mth processing beat is The feature representation vector of the nth clustering result in the (m+1)th processing beat is The threshold is set to Thres, which is set here. Their combined losses are The specific steps for associating radiation source targets are as follows: 1) If only one is found satisfy If the two clustering results represent the same radiation source target, then the nth cluster of the (m+1)th processing beat is directly associated with the nth cluster of the mth processing beat as a radiation source target.

[0031] 2) If multiple results are found satisfy , then it means It may be correlated with multiple clustering results; in this case, choose... The smallest clustering results are associated.

[0032] 3) If all None of them satisfy less than or equal to If the result is not the same as any of the clustering results, then a new radiation source target should be created and reported.

[0033] As another aspect of the present invention, a radiation source target association device based on a depth autoencoder is provided, comprising a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method of any of the above embodiments.

[0034] As another aspect of the present invention, a radiation source target association system based on a depth autoencoder is provided, including the radiation source target association device based on a depth autoencoder as described above.

[0035] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0036] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0037] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A radiation source-target association method based on a deep autoencoder, characterized in that, Includes the following steps: For the clustering results of each processing beat, the sequences of five dimensions, namely frequency, repetition frequency, pulse width, amplitude, and orientation, are extracted. Then, a deep autoencoder based on CNN combined with LSTM network is used to represent the features of the clustering results. The CNN network represents the local features of the clustering results, and the LSTM network represents the temporal features of the clustering results. In the process of realizing radiation source target association in the deep autoencoder inference application, the clustering results of each processing beat are first obtained by using a trained deep autoencoder network to obtain feature representation vectors. Then, based on the joint loss of the feature representation vectors of different processing beats, it is determined whether the two corresponding clustering results belong to the same radiation source target, thereby realizing stable association of radiation source targets.

2. The radiation source target association method based on a deep autoencoder according to claim 1, characterized in that, The training loss function of the deep autoencoder network is a joint loss function constructed by combining mean squared error and cosine similarity loss.

3. The radiation source target association method based on a deep autoencoder according to claim 1, characterized in that, The method of constructing a deep autoencoder based on a CNN combined with an LSTM network to represent the features of the clustering results, where the CNN network represents the local features of the clustering results and the LSTM network represents the temporal features of the clustering results, specifically includes the following sub-steps: Let the clustering result of the nth clustering in the mth processing beat be represented as: ,in This represents the frequency of the k-th pulse in the clustering result. This represents the repetition frequency of the k-th pulse in the clustering result. This represents the pulse width of the k-th pulse in the clustering result. This represents the amplitude of the k-th pulse in the clustering result. This indicates the orientation [Z1] of the k-th pulse in the clustering result. M represents the maximum number of processing cycles, N represents the maximum number of clusters in each processing cycle, and K represents the maximum number of pulses in each cluster result; CNN networks are composed of stacked convolutional blocks consisting of convolution operations, activation function processing, batch normalization, and pooling. The specific operation steps are as follows: ； in, This indicates convolution processing. This indicates activation function processing. This indicates normalization processing. Indicates pooling processing; , This indicates the number of times convolutional blocks are stacked in a CNN network; After constructing the CNN network, an LSTM network is used to further extract the temporal features of the clustering network. The formula is as follows: ； in, This represents the hidden layer of the LSTM network at the last moment of time series processing, serving as the feature representation of the corresponding deep autoencoder input; Then, with As input, the constructed deep autoencoder is used for processing to obtain the feature representation vectors of the clustering results in each processing beat. ,set up ,in express Dimensions These represent the parameters for each dimension.

4. The radiation source target association method based on a deep autoencoder according to claim 2, characterized in that, The training of the deep autoencoder network specifically includes the following sub-steps: The constructed deep autoencoder network is trained by randomly selecting two sample pairs from the training set for processing, and constructing a joint loss based on mean squared error and cosine similarity loss. At the same time, the Adam optimizer is used to train and obtain the parameters of the deep autoencoder network.

5. The radiation source target association method based on a deep autoencoder according to claim 2, characterized in that, The clustering results for each processing beat are used to obtain feature representation vectors through a trained deep autoencoder network. Then, based on the joint loss of the feature representation vectors of different processing beats, it is determined whether the two corresponding clustering results belong to the same radiation source target. This specifically includes the following sub-steps: For clustering results with different processing rhythms, the clustering results are first processed by the trained deep autoencoder to obtain the feature representation of the corresponding clustering results. Then, for the feature representation vectors of different processing rhythms, the joint loss based on mean square error and cosine similarity is calculated respectively, and a threshold is set to determine whether the two clustering results belong to the same radiation source target.

6. The radiation source target association method based on a deep autoencoder according to claim 5, characterized in that, The step of calculating the joint loss based on mean square error and cosine similarity, and setting a threshold, to determine whether the two clustering results belong to the same radiation source target, specifically includes the following sub-steps: Let the feature representation vector of the nth clustering result in the mth processing beat be... The feature representation vector of the nth clustering result in the (m+1)th processing beat is The threshold is set to Thres, and the joint loss is... The sub-steps for determining whether two clustering results belong to the same radiation source target are as follows: 1) If only one is found satisfy If the two clustering results represent the same radiation source target, then the nth cluster of the (m+1)th processing beat is directly associated with the nth cluster of the mth processing beat as a radiation source target; 2) If multiple results are found satisfy , then it means It may be correlated with multiple clustering results; in this case, choose... Associate the smallest cluster results; 3) If all None of them satisfy less than or equal to If the result is not the same as any of the clustering results, then a new radiation source target should be created and reported.

7. A radiation source target association device based on a depth autoencoder, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, A computer program is stored in a readable storage medium, which, when loaded by a processor, executes the method as described in any one of claims 1 to 6.

9. A radiation source-target association system based on a deep autoencoder, characterized in that, Includes the radiation source target association device based on a depth autoencoder as described in claim 7.

Images