Spaceborne SAR Operating Mode Recognition Method and Device Based on Multi-Scale Network

Abstract

The present invention discloses a method and device for identifying the operating mode of spaceborne SAR based on a multi-scale network, including: obtaining a multi-source reconnaissance parameter sequence of a spaceborne synthetic aperture radar reconnaissance signal to be identified, extracting synthetic aperture radar semantic information to obtain a synthetic aperture radar semantic sequence, and performing normalization processing to obtain a normalized synthetic aperture radar semantic sequence; calculating the mean value of the normalized synthetic aperture radar semantic sequence to obtain the mean value of the synthetic aperture radar semantic sequence; extracting an amplitude sequence, performing normalization processing on the amplitude sequence to obtain a normalized amplitude sequence; using a trained spaceborne synthetic aperture radar operating mode recognition network to process the mean value of the synthetic aperture radar semantic sequence and the normalized amplitude sequence to obtain the operating mode corresponding to the multi-source reconnaissance parameter sequence. The present invention can improve the accuracy of identifying the operating mode of spaceborne synthetic aperture radar.

Description

Technical Field

[0001] The present invention belongs to the technical field of synthetic aperture radar reconnaissance signal processing, and particularly relates to a method and device for identifying the working mode of spaceborne SAR based on a multi-scale network. Background Art

[0002] Synthetic Aperture Radar (SAR) is a microwave remote sensing radar, which has the advantages of long distance and wide coverage, and can work all day and all weather in bad weather environments such as clouds and rain. It is widely used in the fields of ocean observation, geological mapping, information collection, etc. SAR is mainly divided into airborne SAR and spaceborne SAR, and spaceborne SAR has become a research hotspot due to its global observation and other advantages. However, when using spaceborne SAR to reconnoiter the other party's scene, the key areas of one's own side will also be correspondingly exposed to the detection of the other party's spaceborne SAR. Therefore, it is necessary to take corresponding measures to disrupt the normal operation of the other party's spaceborne SAR. A typical method is to jam the spaceborne SAR to prevent the other party from obtaining one's own information. Whether the waveform parameters, working mode and other information of the other party can be accurately obtained before jamming the spaceborne SAR will have an important impact on the effectiveness of the jamming. Therefore, carrying out research on the precise processing method of spaceborne SAR reconnaissance signals has an important promoting effect on the efficient implementation of spaceborne SAR jamming.

[0003] With the development of spaceborne SAR systems from early low-resolution and single-mode to high-resolution and multi-mode, etc., multi-mode SAR with higher flexibility has been widely used and deployed, bringing severe challenges to the accurate identification of the working mode of spaceborne SAR. In recent years, many scholars have actively explored the problem of SAR working mode identification. Tang Xiaoming et al. constructed an error sum of squares array according to theory and intercepted data, and used iterative search to infer the SAR working mode and observation range. In addition, a genetic algorithm was used to accelerate the search speed, and a multi-station measurement method was proposed to improve the reliability of the algorithm. Wang Zhetao et al. used a binary tree support vector machine to realize the identification of spaceborne SAR imaging modes, and used a genetic algorithm to optimize the parameters in the model. He Jun et al. proposed a method for identifying SAR working modes based on CGRU-SVM and a one-dimensional convolutional neural network, and used the I / Q data and amplitude of the pulse peak as the inputs of the two networks respectively to achieve the rapid identification of spaceborne SAR working modes. Although the above methods have all studied the problem of SAR working mode identification, they still have the problems of long time consumption, poor generality, and low recognition accuracy in complex scenarios. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a method and device for identifying the working mode of spaceborne SAR based on a multi-scale network. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0005] In a first aspect, the present invention provides a method for identifying the operating mode of a spaceborne SAR based on a multi-scale network, including:

[0006] Obtain a multi-source reconnaissance parameter sequence of the spaceborne synthetic aperture radar reconnaissance signal to be identified;

[0007] Extract the synthetic aperture radar semantic information in the multi-source reconnaissance parameter sequence to obtain a synthetic aperture radar semantic sequence; perform normalization processing on the synthetic aperture radar semantic sequence to obtain a normalized synthetic aperture radar semantic sequence; calculate the mean value of the normalized synthetic aperture radar semantic sequence to obtain the mean value of the synthetic aperture radar semantic sequence;

[0008] Extract the amplitude sequence in the multi-source reconnaissance parameter sequence to obtain an amplitude sequence, perform normalization processing on the amplitude sequence to obtain a normalized amplitude sequence;

[0009] Use the trained spaceborne synthetic aperture radar operating mode recognition network to process the mean value of the synthetic aperture radar semantic sequence and the normalized amplitude sequence to obtain the operating mode corresponding to the multi-source reconnaissance parameter sequence;

[0010] Among them, the trained spaceborne synthetic aperture radar operating mode recognition network uses data of a preset category as a training data set to train the initial spaceborne synthetic aperture radar operating mode recognition network, and performs model selection and hyperparameter tuning on the spaceborne synthetic aperture radar operating mode recognition network using data of a preset category as a validation data set during the training process.

[0011] In a second aspect, the present invention further provides a device for identifying the operating mode of a spaceborne SAR based on a multi-scale network, including a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory complete communication with each other through the communication bus;

[0012] The memory is used to store a computer program;

[0013] The processor is used to implement the method for identifying the operating mode of a spaceborne SAR based on a multi-scale network provided above when executing the program stored on the memory.

[0014] Advantages of the present invention:

[0015] A method and device for identifying the operating mode of spaceborne SAR based on a multi-scale network are provided by the present invention. By introducing synthetic aperture radar semantic information such as duty cycle and range resolution that are tightly coupled with the operating mode of the synthetic aperture radar, the ability of the network to identify different operating modes under non-ideal conditions is improved. The trained spaceborne synthetic aperture radar operating mode recognition network fully fuses the multi-scale information of the mean value of the synthetic aperture radar semantic sequence and the normalized amplitude sequence, improving the discrimination of different operating modes of the spaceborne synthetic aperture radar in the high-dimensional latent feature space, thereby improving the accuracy of identifying the operating mode of the spaceborne synthetic aperture radar. In addition, experiments in complex scenarios with a high proportion of missing pulses and false pulses show that, compared with existing methods for identifying the operating mode of spaceborne synthetic aperture radar, the method for identifying the operating mode of spaceborne synthetic aperture radar based on a knowledge-assisted multi-scale information fusion network has good recognition accuracy and robustness.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0017] Figure 1 is a flowchart of a method for identifying the operating mode of spaceborne SAR based on a multi-scale network provided by an embodiment of the present invention;

[0018] Figure 2 is a schematic diagram of a trained spaceborne synthetic aperture radar operating mode recognition network provided by an embodiment of the present invention;

[0019] Figure 3 is a schematic diagram of the result curves of different methods for identifying the operating mode of spaceborne synthetic aperture radar during the process of increasing the false pulse ratio from 5% to 20% provided by an embodiment of the present invention;

[0020] Figure 4 is a schematic diagram of the result curves of different methods for identifying the operating mode of spaceborne synthetic aperture radar during the process of increasing the missing pulse ratio from 5% to 20% provided by an embodiment of the present invention. Detailed Embodiments

[0021] The present invention will be further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0022] Please refer to Figure 1 , Figure 1 is a flowchart of a method for identifying the operating mode of spaceborne SAR based on a multi-scale network provided by an embodiment of the present invention. A method for identifying the operating mode of spaceborne SAR based on a multi-scale network provided by the present invention includes:

[0023] S101. Obtain a multi-source reconnaissance parameter sequence of a spaceborne synthetic aperture radar reconnaissance signal to be identified.

[0024] Specifically, in this embodiment, the multi-parameter reconnaissance sequence of the on-orbit synthetic aperture radar reconnaissance signal to be recognized can be a parameter sequence obtained by a reconnaissance receiver measuring the parameters of the signal emitted during the operation of the on-orbit synthetic aperture radar.

[0025] S102. Extract the synthetic aperture radar semantic information from the multi-parameter reconnaissance sequence to obtain a synthetic aperture radar semantic sequence; perform normalization processing on the synthetic aperture radar semantic sequence to obtain a normalized synthetic aperture radar semantic sequence; calculate the mean value of the normalized synthetic aperture radar semantic sequence to obtain the mean value of the synthetic aperture radar semantic sequence.

[0026] In this embodiment, the synthetic aperture radar semantic sequence includes a duty cycle sequence and a range resolution sequence.

[0027] Specifically, considering that the range resolution and the duty cycle are closely related to the quality of the on-orbit synthetic aperture radar image, in this embodiment, the range resolution and the duty cycle are used as important feature quantities. Among them, the range resolution is determined by the signal bandwidth (Bandwidth, BW), and the duty cycle is related to the pulse width (Pulse Width, PW) and the pulse repetition interval (Pulse Repetition Interval, PRI). During the actual operation of the on-orbit SAR, the BW, PW, and PRI will be appropriately adjusted according to different mission requirements. Since the range resolution and the duty cycle are features tightly coupled with the characteristics of the on-orbit SAR working mode and can reflect the combined changes of the signal BW, PW, and PRI. Therefore, the range resolution and duty cycle features can be extracted from the multi-parameter reconnaissance sequence received by the reconnaissance receiver to assist in the subsequent identification of the SAR working mode.

[0028] Specifically, for an on-orbit synthetic aperture radar system that emits linear frequency modulation signals, the larger the signal bandwidth, the higher the range resolution ability of the on-orbit synthetic aperture radar. Generally, different working modes of the on-orbit synthetic aperture radar system perform specific imaging tasks by selecting different signal bandwidths, and the on-orbit synthetic aperture radar system selects an appropriate duty cycle according to the imaging characteristics of different working modes; for example, in the scan mode, in order to reduce range ambiguity and reduce the overlap of pulse echoes, a lower duty cycle is generally used; while in the spotlight mode, a higher duty cycle is generally used to improve the signal-to-noise ratio and thus obtain a better synthetic aperture radar image. It can be seen that the range resolution and the duty cycle have a tightly coupled relationship with the working mode of the on-orbit synthetic aperture radar system, and they can be extracted as synthetic aperture radar semantic information to improve the discrimination of the on-orbit synthetic aperture radar working mode.

[0029] Obtain the range resolution sequence and the duty cycle sequence, and their expressions are respectively:

[0030] ;

[0031] ;

[0032] Among them, represents the range resolution sequence, represents the speed of light, represents the signal bandwidth sequence in the multi-source reconnaissance parameter sequence, represents the duty cycle sequence. During the actual working process, the spaceborne synthetic aperture radar selects an appropriate duty cycle according to the imaging characteristics of different working modes, represents the pulse width sequence in the multi-source reconnaissance parameter sequence, represents the pulse repetition period sequence in the multi-source reconnaissance parameter sequence.

[0033] Normalize the range resolution sequence and the duty cycle sequence respectively to obtain the normalized range resolution sequence and the normalized duty cycle sequence , and their expressions are:

[0034] ;

[0035] Among them, represents the normalization threshold of the range resolution sequence, represents the normalization threshold of the duty cycle sequence, and the function has the following expression:

[0036] ;

[0037] Among them, represents or ;

[0038] Calculate the mean of the normalized range resolution sequence and the mean of the normalized duty cycle sequence. Their expressions are respectively:

[0039] ;

[0040] ;

[0041] Among them, represents the mean of the normalized range resolution sequence, represents the number of elements in the normalized range resolution sequence, represents the th element in the normalized range resolution sequence, represents the mean of the normalized duty cycle sequence, represents the th element in the normalized duty cycle sequence, represents the summation operation.

[0042] S103. Extract the amplitude sequence from the multi - element reconnaissance parameter sequence to obtain the amplitude sequence, and perform normalization processing on the amplitude sequence to obtain the normalized amplitude sequence.

[0043] Specifically, in this embodiment, the maximum - minimum normalization method is used to normalize the amplitude sequence to the range of [0, 1] to avoid the influence of different variable dimensions during the processing.

[0044] The expression of the normalized amplitude sequence is:

[0045] ;

[0046] Where, represents the amplitude sequence in the multi - element reconnaissance parameter sequence of the space - borne synthetic aperture radar reconnaissance signal to be recognized, represents the minimum - value function, represents the maximum - value function, represents the normalized amplitude sequence.

[0047] S104. Use the trained space - borne synthetic aperture radar working mode recognition network to process the mean of the synthetic aperture radar semantic sequence and the normalized amplitude sequence to obtain the working mode corresponding to the multi - element reconnaissance parameter sequence.

[0048] Among them, the trained space - borne synthetic aperture radar working mode recognition network is obtained by using the data of the preset category as the training data set to train the initial space - borne synthetic aperture radar working mode recognition network, and using the data of the preset category as the validation data set during the training process to perform model selection and hyperparameter tuning on the space - borne synthetic aperture radar working mode recognition network.

[0049] Specifically, please refer to Figure 2 ., Figure 2 is a schematic diagram of a trained space - borne synthetic aperture radar working mode recognition network provided by an embodiment of the present invention. In this embodiment, the trained space - borne synthetic aperture radar working mode recognition network includes a trained feature embedding network, a trained feature extraction network, and a trained classifier. The trained feature embedding network includes a trained amplitude sequence feature embedding network and a trained synthetic aperture radar semantic feature embedding network; using the trained space - borne synthetic aperture radar working mode recognition network to process the mean of the synthetic aperture radar semantic sequence and the normalized amplitude sequence to obtain the working mode corresponding to the multi - element reconnaissance parameter sequence, including:

[0050] Process the mean value of the synthetic aperture radar semantic sequence using the trained synthetic aperture radar semantic feature embedding network to obtain semantic embedding vectors of different scales; process the normalized amplitude sequence using the trained amplitude sequence feature embedding network, perform patch processing with different patch lengths to obtain patch sequences of different scales, and obtain amplitude sequence embedding vectors of different scales according to the patch sequences of different scales; concatenate the semantic embedding vectors and amplitude sequence embedding vectors of the same scale to obtain concatenated embedding vectors of different scales;

[0051] Process the concatenated embedding vectors of different scales using the trained feature extraction network to obtain deep latent features;

[0052] Classify the deep latent features using the trained classifier to identify the working mode corresponding to the multi-source reconnaissance parameter sequence.

[0053] In this embodiment, the trained synthetic aperture radar semantic feature embedding network includes a first fully connected layer, an activation layer, and a second fully connected layer;

[0054] Process the mean value of the synthetic aperture radar semantic sequence using the synthetic aperture radar semantic feature embedding network, transform the dimension of the mean value of the synthetic aperture radar semantic sequence through the first fully connected layer to obtain a feature vector with expanded dimension, use the activation layer to increase the non-linear representation ability of the feature vector with expanded dimension to obtain an activated feature vector, and transform the dimension of the activated feature vector through the second fully connected layer to obtain a deep semantic embedding vector; rearrange the deep semantic embedding vectors according to different scales to obtain semantic embedding vectors of different scales.

[0055] It should be noted that the fully connected layer is used for feature dimension transformation, and the activation layer is a Tanh activation layer, which is used to increase the non-linear representation ability of the network.

[0056] In this embodiment, the trained amplitude sequence feature embedding network includes a multi-scale processing layer, a first linear mapping layer, a second linear mapping layer, a third linear mapping layer, and a first layer normalization layer; the first linear mapping layer, the second linear mapping layer, and the third linear mapping layer all include a third fully connected layer and a position encoding layer; it should be noted that the parameters of the third fully connected layer and the position encoding layer included in the first linear mapping layer, the second linear mapping layer, and the third linear mapping layer are different;

[0057] Process the normalized amplitude sequence using the multi-scale processing layer to obtain patch sequences of different scales;

[0058] The first linear mapping layer, the second linear mapping layer, and the third linear mapping layer are respectively used to process sequences of patches at different scales, embed them into sequence embedding vectors of patches at different scales through the third fully connected layer, and generate encoded vectors at different scales through the position encoding layer;

[0059] Add the sequence embedding vectors of patches at different scales to the corresponding encoded vectors with the same scale to obtain amplitude vectors at different scales;

[0060] Use the first layer normalization layer to normalize the amplitude vectors at different scales to obtain sequence embedding vectors of amplitude at different scales.

[0061] It should be noted that the multi-scale processing layer is used to process the normalized amplitude sequences into subsequences at multiple scales, the position encoding layer is used to record the distribution of temporal information in the sample sequence, and the layer normalization layer is used to normalize the feature vectors to improve the network convergence speed.

[0062] In this embodiment, the trained feature extraction network includes a first encoder, a first patch sequence merging layer, a second encoder, a second patch sequence merging layer, and a third encoder; the first encoder, the second encoder, and the third encoder all include multiple layers of the second layer normalization layer, a feed-forward layer, and a multi-head attention mechanism layer, and the parameters of the multiple layers of the second layer normalization layer are different. The first patch sequence merging layer and the second patch sequence merging layer both include a merging layer, a third layer normalization layer, and a fourth fully connected layer;

[0063] Use the first encoder to process the cascade embedding vector of the first scale in the cascade embedding vectors at different scales to obtain the deep latent features of the first scale;

[0064] Use the first patch sequence merging layer to process the deep latent features of the first scale to obtain the first merged deep latent features, and add the cascade embedding vector of the second scale in the cascade embedding vectors at different scales to the first merged deep latent features to obtain the first feature to be processed;

[0065] Use the second encoder to process the first feature to be processed to obtain the deep latent features of the second scale;

[0066] Use the second patch sequence merging layer to process the deep latent features of the second scale to obtain the second merged deep latent features, and add the cascade embedding vector of the third scale in the cascade embedding vectors at different scales to the second merged deep latent features to obtain the second feature to be processed;

[0067] The second feature to be processed is processed by a third encoder to obtain deep latent features.

[0068] It should be noted that the first encoder, the second encoder, and the third encoder are all encoders in the Transformer network. The first patch sequence merging layer and the second patch sequence merging layer are used to merge feature vectors, and the multi-head attention mechanism layer is used to calculate the attention weights between different patches in the patch sequence to capture the dependencies within the sequence.

[0069] In this embodiment, the trained classifier includes an activation layer, a fifth fully connected layer, a sixth fully connected layer, and an activation function layer;

[0070] The activation layer is used to process the deep latent features to improve the non-linear expression ability of the network. The fifth fully connected layer and the sixth fully connected layer are used to compress the information of the feature vectors, and the activation function layer represents the feature vectors output by the fully connected layer in the form of the probabilities of the working mode categories.

[0071] In this embodiment, the training process of the trained spaceborne synthetic aperture radar working mode recognition network includes:

[0072] Obtain data of multiple preset categories;

[0073] Extract the synthetic aperture radar semantic information in the data of the preset categories to obtain a training synthetic aperture radar semantic sequence; perform normalization processing on the training synthetic aperture radar semantic sequence to obtain a normalized training synthetic aperture radar semantic sequence; calculate the mean value of the normalized training synthetic aperture radar semantic sequence to obtain the mean value of the training synthetic aperture radar semantic sequence;

[0074] Extract the amplitude sequence in the data of the preset categories to obtain a training amplitude sequence, and perform normalization processing on the training amplitude sequence to obtain a normalized training amplitude sequence;

[0075] The mean values of the training synthetic aperture radar semantic sequences corresponding to the data of the preset categories and the normalized training amplitude sequences are used as samples in the training dataset, and the mean values of the training synthetic aperture radar semantic sequences corresponding to the data of the preset categories and the normalized training amplitude sequences are used as samples in the validation dataset; meanwhile, obtain the true labels of the samples in the training dataset and the validation dataset; and are positive integers greater than 0;

[0076] Input some samples in the training dataset into the Train the on - satellite synthetic aperture radar (SAR) working mode recognition network to be trained for the th time to obtain the prediction results output by the classifier during the

[0077] According to the prediction results output by the classifier during the th training process and the true labels of the samples for training the on - satellite SAR working mode recognition network for the th time, calculate the classification loss, and use it as the classification loss for the th training process;

[0078] According to the classification loss of the th training process, perform backpropagation to update the network parameters of the on - satellite SAR working mode recognition network to be trained for the th time, and obtain the on - satellite SAR working mode recognition network to be trained for the th time; At the same time, every time a preset verification interval value is reached. For example, take 20 training times as the preset verification interval value, input the samples in the verification dataset into the on - satellite SAR working mode recognition network to be trained, obtain the prediction results output by the classifier, and combine with the true labels of the samples in the verification dataset to perform model selection and hyperparameter tuning. Iterate in this way until the training times or the convergence degree meet the preset conditions, and obtain the trained on - satellite SAR working mode recognition network.

[0079] In summary, a method for on - satellite SAR working mode recognition provided by the present invention improves the network's recognition ability for different working modes under non - ideal conditions by introducing SAR semantic information such as duty cycle and range resolution that are tightly coupled with the SAR working mode. The trained on - satellite SAR working mode recognition network fully fuses the mean of the SAR semantic sequence and the multi - scale information of the normalized amplitude sequence, improving the distinguishability of different working modes of the on - satellite SAR in the high - dimensional latent feature space, thereby improving the accuracy of on - satellite SAR working mode recognition. In addition, experiments in complex scenarios with a high proportion of missing pulses and false pulses show that compared with existing on - satellite SAR working mode recognition methods, the on - satellite SAR working mode recognition method based on the knowledge - assisted multi - scale information fusion network has good recognition accuracy and robustness.

[0080] In an optional embodiment of the present invention, the on - satellite SAR working mode recognition of the present invention is implemented through the following process, specifically:

[0081] S201. Obtain the multi - element reconnaissance parameter sequence of the on - satellite SAR reconnaissance signal.

[0082] S202. Extract synthetic aperture radar semantic information such as the range resolution sequence and duty cycle sequence that have a tightly coupled relationship with the spaceborne synthetic aperture radar operating mode from the multi-source reconnaissance parameter sequence to obtain the synthetic aperture radar semantic sequence, and perform normalization processing on the synthetic aperture radar semantic sequence to obtain the normalized synthetic aperture radar semantic sequence, which are the normalized range resolution sequence and the normalized duty cycle sequence respectively.

[0083] S203. Calculate the means of the normalized range resolution sequence and the normalized duty cycle sequence to obtain the means of the synthetic aperture radar semantic sequence, which are the mean of the range resolution sequence and the mean of the duty cycle sequence respectively.

[0084] S204. Extract the amplitude sequence from the multi-source reconnaissance parameter sequence, and perform normalization processing on the amplitude sequence using the maximum-minimum normalization method to obtain the normalized amplitude sequence.

[0085] S205. Construct multiple samples, each sample including the normalized amplitude sequence and the mean of the synthetic aperture radar semantic sequence; divide the multiple samples into a training data set, a validation data set, and a test data set; at the same time, label the true labels of the operating modes corresponding to the samples in the training data set and the validation data set.

[0086] Optionally, the number of samples in the training data set, the validation data set, and the test data set can be determined according to the actual situation, and the present invention does not make specific limitations here.

[0087] S206. Construct a spaceborne synthetic aperture radar operating mode recognition network.

[0088] Please continue to refer to Figure 2 , the spaceborne synthetic aperture radar operating mode recognition network includes a feature embedding network, a feature extraction network, and a classifier.

[0089] Among them, the feature embedding network includes an amplitude sequence feature embedding network and a synthetic aperture radar semantic feature embedding network; the synthetic aperture radar semantic feature embedding network includes a first fully connected layer, an activation layer, and a second fully connected layer; the amplitude sequence feature embedding network includes a multi-scale processing layer, a first linear mapping layer, a second linear mapping layer, a third linear mapping layer, and a first layer normalization layer; the first linear mapping layer, the second linear mapping layer, and the third linear mapping layer all include a third fully connected layer and a position encoding layer; it should be noted that the parameters of the third fully connected layer and the position encoding layer included in the first linear mapping layer, the second linear mapping layer, and the third linear mapping layer are different.

[0090] The feature extraction network includes a first encoder, a first patch sequence merging layer, a second encoder, a second patch sequence merging layer, and a third encoder; the first encoder, the second encoder, and the third encoder each include multiple layers of second layer normalization layers, feed-forward layers, and multi-head attention mechanism layers, and the parameters of the multiple layers of second layer normalization layers are different. The first patch sequence merging layer and the second patch sequence merging layer each include a merging layer, a third layer normalization layer, and a fourth fully connected layer.

[0091] The classifier includes an activation layer, a fifth fully connected layer, a sixth fully connected layer, and an activation function layer; the classifier is used to establish a mapping relationship between the deep latent features and the corresponding spaceborne synthetic aperture radar working mode categories.

[0092] It should be noted that at this time, the network parameters of the spaceborne synthetic aperture radar working mode recognition network are not determined.

[0093] S207. Initialize the spaceborne synthetic aperture radar working mode recognition network, input the training data set and the validation data set into the initial spaceborne synthetic aperture radar working mode recognition network. The samples in the training data set train the spaceborne synthetic aperture radar working mode recognition network, and iteratively optimize the network parameters under the guidance of the loss function until convergence. The samples in the validation data set are used for model selection and hyperparameter tuning during the training process to obtain a trained spaceborne synthetic aperture radar working mode recognition network.

[0094] S208. Input the samples in the test data set into the trained spaceborne synthetic aperture radar working mode recognition network to identify the working mode categories corresponding to the samples in the test data set.

[0095] S2081. Input the samples in the test data set into the trained feature embedding network to respectively output semantic embedding vectors of different scales and amplitude sequence embedding vectors of different scales.

[0096] S2082. Concatenate the semantic embedding vectors and amplitude sequence embedding vectors of the same scale to obtain concatenated embedding vectors of different scales.

[0097] S2083. Input the concatenated embedding vectors of different scales into the trained feature extraction network for processing to obtain deep latent features.

[0098] S2084. Input the deep latent features into the trained classifier for processing to obtain the working mode corresponding to the samples in the test data set.

[0099] In an optional embodiment of the present invention, the effect of the spaceborne SAR working mode recognition method based on a multi-scale network provided in the above embodiment is verified through a simulation experiment, specifically:

[0100] I. Simulation Conditions

[0101] The simulation experiment conditions of this embodiment: Simulate the data of the strip, spotlight, sliding spotlight, and scanning working modes of the spaceborne synthetic aperture radar. The value ranges of the simulation parameters of different working modes of the spaceborne synthetic aperture radar are shown in Table 1. Among them, 1000 samples are generated for each working mode, and the total number of samples is 4000. Randomly select 70% of the total number of samples to construct the training data set, 10% to construct the validation data set, and 20% to construct the test data set. At the same time, in order to test the adaptability of the method proposed in the present invention in a complex environment containing missing pulses and false pulses, the proportions of missing pulses and false pulses are respectively set between 0% and 20%.

[0102] Table 1 Value ranges of simulation parameters of different working modes of spaceborne synthetic aperture radar

[0103]

[0104] II. Simulation Content and Result Analysis

[0105] Please refer to Table 2. The batch size is the number of samples used to update the model parameters in each iteration process. The epoch is the number of times to completely train the model using all training samples. The initial learning rate is the initial learning rate. The optimizer is the optimizer. The V_dim is the dimension of the hidden layer feature vector. The N_heads is the number of heads in the multi-head attention mechanism. d ff The dimension of the fully connected layer is the dimension of the fully connected layer. The layers is the number of Transformer encoders. Randomly select 70% of all samples for training, 10% for validation, and the remaining 20% for testing. Among them, the validation data set is used to verify the performance of the spaceborne synthetic aperture radar working mode recognition network trained in some iteration times, and the test data set is used to test the performance of the overall trained spaceborne synthetic aperture radar working mode recognition network.

[0106] Table 2 Parameters of the spaceborne synthetic aperture radar working mode recognition network

[0107]

[0108] The method proposed in the present invention tests the performance under different proportions of missing pulses and false pulses. The results show that as the proportions of false pulses and missing pulses gradually increase, the algorithm performance gradually decreases. This is because the existence of false pulses and missing pulses will disrupt the timing rules and joint feature rules of the data in each working mode, and the higher the proportion, the more serious the disruption of the data rules, which in turn makes it difficult for the network to extract highly discriminative features. It should be noted that the method for identifying the working mode of spaceborne SAR based on a multi-scale network in the present invention can still maintain the highest recognition accuracy when the proportion of missing pulses or false pulses is as high as 20%. This is due to the introduction of the semantic information of synthetic aperture radar that has a tightly coupled relationship with the working mode of spaceborne synthetic aperture radar, which weakens the influence of outliers on the network; and the adoption of multi-scale information fusion, which effectively improves the recognition accuracy of the working mode of spaceborne synthetic aperture radar in non-ideal situations containing missing pulses and false pulses.

[0109] Please refer to Figure 3 , Figure 3 which is a schematic diagram of the result curves of different methods for identifying the working mode of spaceborne synthetic aperture radar during the process of increasing the proportion of false pulses from 5% to 20% provided by an embodiment of the present invention. Figure 3 is the accuracy curve. The abscissa represents the proportion of added false pulses, and the ordinate represents the accuracy. The curve with squares is the experimental result of the method proposed in the present invention, the curve with circles is the experimental result of the CNN method, the curve with pentagrams is the experimental result of the Transformer method, and the curve with triangles is the experimental result of the TCN method. It can be seen that as the proportion of false pulses gradually increases, the experimental results of all methods show varying degrees of decline. However, the method proposed in the present invention always maintains the highest accuracy during this process. This result indicates that the method proposed in the present invention has good adaptability in non-ideal situations.

[0110] Please refer to Figure 4 , Figure 4 which is a schematic diagram of the result curves of different methods for identifying the working mode of spaceborne synthetic aperture radar during the process of increasing the proportion of missing pulses from 5% to 20% provided by an embodiment of the present invention. Figure 4It is the accuracy curve. The abscissa represents the proportion of added missing pulses, and the ordinate represents the accuracy. The curve with squares in the figure is the experimental result of the method proposed in the present invention, the curve with circles is the experimental result of the CNN method, the curve with pentagrams is the experimental result of the Transformer method, and the curve with triangles is the experimental result of the TCN method. It can be seen that as the proportion of missing pulses gradually increases, the experimental results of all methods show varying degrees of decline. However, the method proposed in the present invention always maintains the highest accuracy during this process. This result indicates that the method proposed in the present invention has good adaptability in non-ideal situations.

[0111] The method proposed in the present invention introduces synthetic aperture radar semantic information such as range resolution and duty cycle, which has a tight coupling relationship with the spaceborne synthetic aperture radar working mode, into the spaceborne synthetic aperture radar working mode recognition task according to the expert knowledge of synthetic aperture radar, enriching the representation of the working mode. On this basis, aiming at the problem of low accuracy in recognizing the working mode of spaceborne synthetic aperture radar in complex non-ideal scenarios with a high proportion of missing pulses and false pulses, a spaceborne synthetic aperture radar working mode recognition network is constructed. Based on cascading multi-scale amplitude sequence embedding vectors and semantic embedding vectors, the network uses a feature extraction module to fully fuse the cascaded embedding vectors of different scales, effectively improving the discrimination of different working modes of spaceborne synthetic aperture radar in the high-dimensional latent feature space, and thus effectively improving the accuracy of recognizing the working mode of spaceborne synthetic aperture radar.

[0112] Based on the same inventive concept, the present invention also provides a spaceborne SAR working mode recognition device based on a multi-scale network, including a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory complete mutual communication through the communication bus;

[0113] The memory is used for storing computer programs;

[0114] The processor is used to implement the above-provided spaceborne SAR working mode recognition method based on a multi-scale network when executing the programs stored on the memory.

[0115] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variants are intended to cover non-exclusive inclusion, so that an article or device comprising a series of elements includes not only those elements but also other elements not expressly listed. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the article or device comprising the element. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "upper", "lower", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation on the present invention.

[0116] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0117] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for identifying spaceborne SAR working modes based on multi-scale networks, characterized in that: include: Acquire a multi-element reconnaissance parameter sequence of a spaceborne synthetic aperture radar reconnaissance signal to be identified; Extracting synthetic aperture radar semantic information from the multivariate reconnaissance parameter sequence to obtain a synthetic aperture radar semantic sequence; normalizing the synthetic aperture radar semantic sequence to obtain a normalized synthetic aperture radar semantic sequence; calculating a mean of the normalized synthetic aperture radar semantic sequence to obtain a mean of the synthetic aperture radar semantic sequence; Extracting the amplitude sequence in the multivariate reconnaissance parameter sequence to obtain an amplitude sequence, and normalizing the amplitude sequence to obtain a normalized amplitude sequence; Using a trained spaceborne synthetic aperture radar working mode recognition network, the mean of the synthetic aperture radar semantic sequence and the normalized amplitude sequence are processed to obtain the working mode corresponding to the multivariate reconnaissance parameter sequence; The trained spaceborne synthetic aperture radar working mode recognition network uses preset category data as a training data set to train the initial spaceborne synthetic aperture radar working mode recognition network, and during the training process, the preset category data is used as a verification data set to perform model selection and hyperparameter tuning on the spaceborne synthetic aperture radar working mode recognition network.

2. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 1, characterized in that: The trained spaceborne synthetic aperture radar working mode recognition network includes a trained feature embedding network, a trained feature extraction network and a trained classifier, and the trained feature embedding network includes a trained amplitude sequence feature embedding network and a trained synthetic aperture radar semantic feature embedding network; The method uses a trained spaceborne synthetic aperture radar working mode recognition network to process the mean of the synthetic aperture radar semantic sequence and the normalized amplitude sequence to obtain the working mode corresponding to the multivariate reconnaissance parameter sequence, including: The trained synthetic aperture radar semantic feature embedding network is used to process the mean of the synthetic aperture radar semantic sequence to obtain semantic embedding vectors of different scales; the trained amplitude sequence feature embedding network is used to process the normalized amplitude sequence, and segmentation block processing is performed with different segmentation block lengths to obtain segmentation block sequences of different scales, and amplitude sequence embedding vectors of different scales are obtained according to the segmentation block sequences of different scales; the semantic embedding vector and the amplitude sequence embedding vector of the same scale are cascaded to obtain cascaded embedding vectors of different scales; Using the trained feature extraction network to extract features from the cascade embedding vectors of different scales to obtain deep hidden features; The trained classifier is used to classify the deep latent features and identify the working mode corresponding to the multivariate reconnaissance parameter sequence.

3. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 2, characterized in that: The trained synthetic aperture radar semantic feature embedding network includes a first fully connected layer, an activation layer, and a second fully connected layer; The synthetic aperture radar semantic feature embedding network is used to process the mean of the synthetic aperture radar semantic sequence, the first fully connected layer is used to transform the dimension of the mean of the synthetic aperture radar semantic sequence to obtain a dimensionally expanded feature vector, the activation layer is used to increase the nonlinear representation ability of the dimensionally expanded feature vector to obtain an activated feature vector, the second fully connected layer is used to transform the dimension of the activated feature vector to obtain a deep semantic embedding vector; the deep semantic embedding vector is rearranged according to different scales to obtain semantic embedding vectors of different scales.

4. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 2, characterized in that: The trained amplitude sequence feature embedding network includes a multi-scale processing layer, a first linear mapping layer, a second linear mapping layer, a third linear mapping layer and a first normalization layer; the first linear mapping layer, the second linear mapping layer and the third linear mapping layer all include a third fully connected layer and a position encoding layer; The normalized amplitude sequence is processed by the multi-scale processing layer to obtain a sequence of segmented blocks of different scales; The first linear mapping layer, the second linear mapping layer and the third linear mapping layer are used to process the segmented block sequences of different scales respectively, and the segmented block sequences of different scales are embedded into embedding vectors of the segmented block sequences of different scales through the third fully connected layer, and the encoding vectors of different scales are generated through the position encoding layer; Adding the embedding vectors of the segmented block sequences of different scales to the corresponding encoding vectors of the same scale to obtain amplitude vectors of different scales; The first normalization layer is used to normalize the amplitude vectors of different scales to obtain amplitude sequence embedding vectors of different scales.

5. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 2, characterized in that: The trained feature extraction network includes a first encoder, a first segmentation block sequence merging layer, a second encoder, a second segmentation block sequence merging layer and a third encoder; the first encoder, the second encoder and the third encoder all include a multi-layer second normalization layer, a feedforward layer and a multi-head attention mechanism layer, and the first segmentation block sequence merging layer and the second segmentation block sequence merging layer all include a merging layer, a third normalization layer and a fourth fully connected layer; Using the first encoder to process the concatenated embedding vector of a first scale among the concatenated embedding vectors of different scales to obtain a deep implicit feature of the first scale; Using the first segmentation block sequence merging layer to process the deep implicit features of the first scale to obtain a first merged deep implicit feature, adding the cascade embedding vector of the second scale in the cascade embedding vectors of different scales to the first merged deep implicit feature to obtain a first feature to be processed; Using the second encoder to process the first feature to be processed to obtain a deep latent feature of a second scale; The second segmentation block sequence merging layer is used to process the second scale deep implicit features to obtain second merged deep implicit features, and the cascade embedding vector of the third scale in the cascade embedding vectors of different scales is added to the second merged deep implicit features to obtain a second feature to be processed; The third encoder is used to process the second feature to be processed to obtain a deep latent feature.

6. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 1, characterized in that: The synthetic aperture radar semantic sequence includes a duty cycle sequence and a range resolution sequence.

7. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 6, characterized in that: The step of extracting the synthetic aperture radar semantic information from the multivariate reconnaissance parameter sequence to obtain a synthetic aperture radar semantic sequence; and normalizing the synthetic aperture radar semantic sequence to obtain a normalized synthetic aperture radar semantic sequence; Calculating the mean of the normalized synthetic aperture radar semantic sequence to obtain the mean of the synthetic aperture radar semantic sequence includes: Get the distance resolution sequence and duty cycle sequence, their expressions are: ； ； in, represents the range resolution sequence, represents the speed of light, represents the signal bandwidth sequence in the multivariate reconnaissance parameter sequence, represents the duty cycle sequence, represents the pulse width sequence in the multivariate reconnaissance parameter sequence, represents the pulse repetition period sequence in the multivariate reconnaissance parameter sequence; The distance resolution sequence and the duty cycle sequence are normalized to obtain a normalized distance resolution sequence. and the normalized duty cycle sequence , whose expression is: ； in, represents the normalized threshold of the distance resolution sequence, represents the normalized threshold of the duty cycle sequence, function The expression is: ； in, express or ; The mean of the normalized distance resolution sequence and the mean of the normalized duty cycle sequence are calculated, and their expressions are: ； ； in, represents the mean of the normalized distance resolution sequence, represents the number of elements in the normalized distance resolution sequence, represents the normalized distance resolution sequence elements, represents the mean of the normalized duty cycle sequence, represents the first in the normalized duty cycle sequence elements, Represents a sum operation.

8. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 1, characterized in that: The expression of the normalized amplitude sequence is: ； in, represents the amplitude sequence in the multivariate reconnaissance parameter sequence, represents the minimum value function, represents the maximum value function, Represents the normalized amplitude sequence.

9. The method for identifying spaceborne SAR working modes based on multi-scale networks according to claim 1, characterized in that: The training process of the trained spaceborne synthetic aperture radar working mode recognition network includes: Acquire data of a plurality of preset categories; Extracting synthetic aperture radar semantic information from the preset category of data to obtain a training synthetic aperture radar semantic sequence; normalizing the training synthetic aperture radar semantic sequence to obtain a normalized training synthetic aperture radar semantic sequence; calculating a mean of the normalized training synthetic aperture radar semantic sequence to obtain a mean of the training synthetic aperture radar semantic sequence; Extracting an amplitude sequence from the preset category of data to obtain a training amplitude sequence, and normalizing the training amplitude sequence to obtain a normalized training amplitude sequence; Will The mean of the training synthetic aperture radar semantic sequence corresponding to the preset category of data and the normalized training amplitude sequence are used as samples in the training data set. The mean of the training synthetic aperture radar semantic sequence corresponding to the preset category of data and the normalized training amplitude sequence are used as samples in the verification data set; at the same time, the true labels of the samples in the training data set and the verification data set are obtained; and is a positive integer greater than 0; Input some samples in the training data set into the The spaceborne synthetic aperture radar working mode recognition network to be trained is trained to obtain the first The prediction results output by the classifier during the training process; According to The prediction results output by the classifier during the training process are consistent with the The real labels of the samples of the spaceborne synthetic aperture radar working mode recognition network to be trained are used to calculate the classification loss and serve as the first The classification loss of the training process; According to The classification loss of the first training process is back-propagated to update the The network parameters of the spaceborne synthetic aperture radar working mode recognition network to be trained are obtained. The satellite-borne synthetic aperture radar working mode recognition network to be trained is inputted with the samples in the verification data set into the satellite-borne synthetic aperture radar working mode recognition network to be trained at each preset verification interval value, and the prediction result output by the classifier is obtained. The model selection and hyperparameter tuning are performed in combination with the real labels of the samples in the verification data set, and the training is iterated until the number of training times or the degree of convergence meets the preset conditions, thereby obtaining the trained satellite-borne synthetic aperture radar working mode recognition network.

10. A spaceborne SAR working mode recognition device based on a multi-scale network, comprising a processor, a communication interface, a memory and a communication bus, characterized in that: The processor, the communication interface and the memory communicate with each other via the communication bus; The memory is used to store computer programs; The processor, when used to execute the program stored in the memory, implements the method according to any one of claims 1 to 9.

Images