A Method for Suppressing Sea Clutter in High Sea States Based on Deep Learning

Abstract

The present invention provides a method for suppressing high-sea-state sea clutter based on deep learning, which relates to the technical field of ship radars and includes: obtaining a measured sea clutter data set; obtaining a simulated sea clutter data set and a measured high-sea-state sea clutter data set according to the measured sea clutter data set, and inputting them into a preset high-sea-state sea clutter generation neural network model for adversarial training; then generating a one-dimensional high-sea-state sea clutter data set and a two-dimensional high-sea-state sea clutter data set and inputting them into a once-optimized high-sea-state sea clutter generation neural network model for secondary adversarial training; using the twice-optimized high-sea-state sea clutter generation neural network model to suppress the high-sea-state sea clutter data to be suppressed; the method provided by the present invention simultaneously uses measured and simulated sea clutter data to model it, and suppresses high-sea-state sea clutter in one dimension and two dimensions respectively, which can improve the generalization performance and suppression performance of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of marine radars, and more particularly, to a method for suppressing sea clutter in high sea states based on deep learning. Background Art

[0002] As a key device for target detection in current water transportation, marine radars can effectively help navigating ships avoid obstacles and collisions without being restricted by vision. However, the echo signals received by marine radars are severely interfered by sea clutter in high sea states, resulting in a serious decline in performance. It is very difficult for civilian ships to detect marine obstacles and other ships in high sea states, and navigation safety cannot be guaranteed, which may cause very serious casualties and economic losses. At the same time, if an emergency occurs, it is also very difficult for marine radars to locate the distressed ships for rapid maritime search and rescue. Therefore, it is very important to suppress the interference of sea clutter in radar echoes under high sea state conditions, which has great practical significance in both civilian and military fields.

[0003] There are mainly the following two problems in the method for suppressing sea clutter in high sea states: one is that in the actual measurement of high sea state sea clutter data, it is difficult to collect high sea state sea clutter data due to limitations such as wide sea areas and high costs; the other is that in the establishment of high sea state sea clutter models, when the sea state is low, the sea surface background is relatively uniform, and good results can be obtained by using traditional signal processing-based clutter distribution models for sea clutter suppression. However, in the strong sea clutter background generated in high sea states, the sea surface clutter distribution has complex randomness and mutation, and it becomes very difficult to model the clutter and set the model parameters, resulting in poor sea clutter suppression effect.

[0004] The current prior art discloses an intelligent clutter suppression method for sea surveillance radar, which is applied to the field of radar target detection. Aiming at the problem that the existing sea clutter suppression technology is difficult to cope with the complex and changeable marine environment; the method in the prior art constructs two mirror-symmetrical generative adversarial networks. One network learns the mapping from the original clutter data to the data after clutter suppression, and the other network learns the mapping from the data after clutter suppression to the original clutter data in reverse. The two types of data sets are respectively input into the two GANs, and through the mutual constraints of the two sets of generators and discriminators and the dynamic discrimination of the radar data before and after clutter suppression, a clutter suppression network with the function of sea clutter suppression is finally obtained; the prior art only constructs a sea clutter data set based on the composite K distribution model, lacking measured sea clutter data, which may lead to insufficient generalization performance of the suppression model; due to the lack of high sea state sea clutter data, the sea clutter suppression method based on the deep learning model is prone to overfitting problems. For complex tasks with limited sample size but requiring powerful models, the loss on the training set is small, while the loss on the validation set or test set is large; most of the existing deep learning-based sea clutter suppression methods perform sea clutter suppression on two-dimensional image data, and the suppression performance of the model is insufficient; therefore, the current prior art has the problem of poor suppression performance for high sea state sea clutter due to insufficient complex sea state samples and insufficient model generalization performance. Summary of the Invention

[0005] To overcome the above-mentioned defect of poor suppression performance of high sea state sea clutter in the prior art, the present invention provides a high sea state sea clutter suppression method based on deep learning, which can significantly improve the suppression performance of the model for high sea state sea clutter.

[0006] To solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A high sea state sea clutter suppression method based on deep learning includes the following steps:

[0008] S1: Obtain a measured sea clutter data set;

[0009] S2: According to the measured sea clutter data set, obtain a simulated sea clutter data set and a measured high sea state sea clutter data set;

[0010] S3: Input the simulated sea clutter data set and the measured high sea state sea clutter data set into a preset high sea state sea clutter generation neural network model for adversarial training to obtain a once-optimized high sea state sea clutter generation neural network model;

[0011] S4: Use the once-optimized high sea state sea clutter generation neural network model to obtain a one-dimensional high sea state sea clutter data set and a two-dimensional high sea state sea clutter data set;

[0012] S5: Input the one-dimensional high-sea-state sea clutter dataset and the two-dimensional high-sea-state sea clutter dataset into the once-optimized high-sea-state sea clutter generation neural network model for adversarial training to obtain a twice-optimized high-sea-state sea clutter generation neural network model;

[0013] S6: Obtain the high-sea-state sea clutter data to be suppressed, and use the twice-optimized high-sea-state sea clutter generation neural network model to suppress the high-sea-state sea clutter data to be suppressed.

[0014] Preferably, in step S2, the specific method for obtaining the simulated sea clutter dataset and the measured high-sea-state sea clutter dataset according to the measured sea clutter dataset is as follows:

[0015] S2.1: Divide the measured sea clutter dataset D into a measured low-sea-state sea clutter dataset D L and a measured high-sea-state sea clutter dataset D H ;

[0016] S2.2: Superimpose simulated sea clutter of several distributions on the measured low-sea-state sea clutter dataset D L to obtain a simulated sea clutter dataset D S .

[0017] Preferably, in step S3, the preset high-sea-state sea clutter generation neural network model is specifically:

[0018] The preset high-sea-state sea clutter generation neural network model includes a high-sea-state sea clutter generation network G and a high-sea-state sea clutter discrimination network D;

[0019] The high-sea-state sea clutter generation network G includes a dimension initialization module, a sea clutter encoding module, several deformable attention conversion modules, and a sea clutter decoding module connected in sequence;

[0020] The dimension initialization module includes a first fully connected layer and a first Reshape layer connected in sequence;

[0021] The sea clutter encoding module includes a first convolutional layer, a first batch normalization layer, a first activation layer, and a downsampling layer connected in sequence;

[0022] Each deformable attention conversion module includes a first deformable convolutional layer, a second batch normalization layer, a second activation layer, an attention layer, a weighted product point, and a first residual addition point; the output end of the second activation layer is also connected to the input end of the weighted product point, and the input end of the first deformable convolutional layer is also connected to the input end of the first residual addition point;

[0023] The sea clutter decoding module includes an upsampling layer, a third batch normalization layer, a third activation layer, a second convolutional layer, and a fourth activation layer connected in sequence.

[0024] The high sea state sea clutter discrimination network D includes a plurality of sea clutter downsampling modules, a plurality of spectral normalization convolutional modules, a plurality of deformable convolutional modules, and a sea clutter classification module connected in sequence;

[0025] Each of the sea clutter downsampling modules includes a third convolutional layer, a fourth batch normalization layer, a fifth activation layer, and a second residual addition point connected in sequence; and the input end of the third convolutional layer is also connected to the input end of the second residual addition point;

[0026] The spectral normalization convolutional module includes a fourth convolutional layer, a spectral normalization layer, and a sixth activation layer connected in sequence;

[0027] The deformable convolutional module includes a second deformable convolutional layer and a seventh activation layer connected in sequence;

[0028] The sea clutter classification module includes a second Reshape layer, a second fully connected layer, and an eighth activation layer connected in sequence.

[0029] Preferably, in step S3, the simulated sea clutter data set and the measured high sea state sea clutter data set are input into a preset high sea state sea clutter generation neural network model for adversarial training to obtain a once-optimized high sea state sea clutter generation neural network model. The specific method is as follows:

[0030] S3.1: Use the union of the simulated sea clutter data set and the measured high sea state sea clutter data set as a one-dimensional real sample set X real , and perform time-frequency spectrum conversion on the one-dimensional real sample set X real to obtain a two-dimensional real sample set Z real ;

[0031] S3.2: Respectively construct a one-dimensional noise sample set X noise that satisfies the Gaussian distribution and a two-dimensional noise sample set Z noise that satisfies the Gaussian distribution;

[0032] S3.3: Respectively set the first network parameter θ G1 and the second network parameter θ G2 and assign them to the high sea state sea clutter generation network G to obtain a one-dimensional generation network G1 and a two-dimensional generation network G2. Respectively set the third network parameter θ D1 and the fourth network parameter θ D2 and assign them to the high sea state sea clutter discrimination network D to obtain a one-dimensional discrimination network D1 and a two-dimensional discrimination network D2;

[0033] S3.4: For the one-dimensional noise sample set Xnoise Perform random sampling to obtain a one-dimensional sampling noise sample set And input it into the one-dimensional generation network G1 to obtain a one-dimensional generated sample set And the corresponding one-dimensional generated sample label set Set the element values in the one-dimensional generated sample label set To 0; perform random sampling on the one-dimensional real sample set X real To obtain a one-dimensional sampled real sample set And the corresponding one-dimensional real sample label set Set the element values in the one-dimensional real sample label set To 1;

[0034] Perform random sampling on the two-dimensional noise sample set Z noise To obtain a two-dimensional sampled noise sample set And input it into the two-dimensional generation network G2 to obtain a two-dimensional generated sample set And the corresponding two-dimensional generated sample label set Set the element values in the two-dimensional generated sample label set To 0; perform random sampling on the two-dimensional real sample set Z real To obtain a two-dimensional sampled real sample set And the corresponding two-dimensional real sample label set Set the element values in the two-dimensional real sample label set To 1;

[0035] Where b is the number of sampled samples;

[0036] S3.5: Use the one-dimensional generated sample set And the one-dimensional sampled real sample set To train the one-dimensional discriminant network D1, and set the first discriminant loss function to optimize the one-dimensional discriminant network D1 to obtain the updated third network parameter θ′ D1 And the optimized one-dimensional discriminant network D′1;

[0037] Use the two-dimensional generated sample set And the two-dimensional sampled real sample set To train the two-dimensional discriminant network D2, and set the second discriminant loss function to optimize the two-dimensional discriminant network D2 to obtain the updated fourth network parameter θ′ D2 And the optimized two-dimensional discriminant network D′2;

[0038] S3.6: Perform random sampling on the one-dimensional noise sample set X noise Again to obtain a one-dimensional secondary sampled noise sample set The one-dimensional secondary sampled noise sample set Input into the one-dimensional generation network G1, and output a one-dimensional quadratic generation sample set and the corresponding one-dimensional quadratic generation sample label set Set the element values in the one-dimensional quadratic generation sample label set to 1;

[0039] Perform random sampling on the two-dimensional noise sample set Z noise again to obtain a two-dimensional quadratic sampling noise sample set Input the two-dimensional quadratic sampling noise sample set into the two-dimensional generation network G2, and output a two-dimensional quadratic generation sample set and the corresponding two-dimensional quadratic generation sample label set Set the element values in the two-dimensional quadratic generation sample label set to 1;

[0040] S3.7: Use the one-dimensional quadratic generation sample set to train the one-dimensional generation network G1, and set the first generation loss function to optimize the one-dimensional generation network G1 to obtain the updated first network parameter θ′ G1 and the optimized one-dimensional generation network G′1;

[0041] Use the two-dimensional quadratic generation sample set to train the two-dimensional generation network G2, and set the second generation loss function to optimize the two-dimensional generation network G2 to obtain the updated second network parameter θ′ G2 and the optimized two-dimensional generation network G′2;

[0042] S3.8: Use the optimized one-dimensional generation network G′1, the optimized one-dimensional discriminant network D′1, the optimized two-dimensional generation network G′2, and the optimized two-dimensional discriminant network D′2 to obtain a once-optimized high sea state sea clutter generation neural network model.

[0043] Preferably, in the step S4, use the once-optimized high sea state sea clutter generation neural network model to obtain a one-dimensional high sea state sea clutter data set and a two-dimensional high sea state sea clutter data set. The specific method is as follows:

[0044] Input the one-dimensional noise sample set X noise into the optimized one-dimensional generation network G′1 to obtain a one-dimensional high sea state sea clutter data set D A1 ;

[0045] Input the two-dimensional noise sample set Z noise into the optimized two-dimensional generation network G′2 to obtain a two-dimensional high sea state sea clutter data set D A2 .

[0046] Preferably, in step S5, the one-dimensional high sea state sea clutter data set and the two-dimensional high sea state sea clutter data set are input into the once-optimized high sea state sea clutter generation neural network model for adversarial training to obtain a twice-optimized high sea state sea clutter generation neural network model. The specific method is as follows:

[0047] S5.1: Perform time-frequency spectrum conversion on the measured low sea state sea clutter data set D L to obtain a two-dimensional low sea state sea clutter sample set SZ real ;

[0048] S5.2: Assign the updated first network parameter θ′ G1 and the updated second network parameter θ′ G2 to the high sea state sea clutter generation network G to obtain a one-dimensional secondary generation network SG1 and a two-dimensional secondary generation network SG2. Assign the updated third network parameter θ′ D1 and the fourth network parameter θ′ D2 to the high sea state sea clutter discrimination network D to obtain a one-dimensional secondary discrimination network SD1 and a two-dimensional secondary discrimination network SD2;

[0049] S5.3: Randomly sample the one-dimensional high sea state sea clutter data set D A1 to obtain one-dimensional sea clutter samples Input the one-dimensional sea clutter samples into the one-dimensional secondary generation network SG1 to obtain one-dimensional suppression samples and the corresponding one-dimensional suppression sample label set Set the element values in the one-dimensional suppression sample label set to 0; Randomly sample the measured low sea state sea clutter data set D L to obtain one-dimensional low sea state samples and the corresponding one-dimensional low sea state sample label set Set the element values in the one-dimensional low sea state sample label set to 1;

[0050] Randomly sample the two-dimensional high sea state sea clutter data set D A2 to obtain two-dimensional sea clutter samples Input the two-dimensional sea clutter samples into the two-dimensional secondary generation network SG2 to obtain two-dimensional suppression samples and the corresponding two-dimensional suppression sample label set Set the element values in the two-dimensional suppression sample label set to 0; Randomly sample the two-dimensional low sea state sea clutter sample set SZ real to obtain two-dimensional low sea state samples and the corresponding two-dimensional low sea state sample label set Set the element values in the two-dimensional low sea state sample label set to 1;

[0051] where r is the number of sampled samples;

[0052] S5.4: Use the one-dimensional suppressed samples and one-dimensional low sea state samples to train the one-dimensional quadratic discriminant network SD1, set the third discriminant loss function, optimize the one-dimensional quadratic discriminant network SD1, and obtain the third network parameter θ″ after the second update D1 and the optimized one-dimensional quadratic discriminant network SD′1;

[0053] Use the two-dimensional suppressed samples and two-dimensional low sea state samples to train the two-dimensional quadratic discriminant network SD2, set the fourth discriminant loss function, optimize the two-dimensional quadratic discriminant network SD2, and obtain the fourth network parameter θ″ after the second update D2 and the optimized two-dimensional quadratic discriminant network SD′2;

[0054] S5.5: Randomly sample the one-dimensional high sea state sea clutter dataset D A1 again to obtain one-dimensional secondarily sampled sea clutter samples Input the one-dimensional secondarily sampled sea clutter samples into the one-dimensional quadratic generation network SG1 to obtain one-dimensional secondarily suppressed samples and the corresponding one-dimensional secondarily suppressed sample label set Set the element values in the one-dimensional secondarily suppressed sample label set to 1;

[0055] Randomly sample the two-dimensional high sea state sea clutter dataset D A2 again to obtain two-dimensional secondarily sampled sea clutter samples Input the two-dimensional secondarily sampled sea clutter samples into the two-dimensional quadratic generation network SG2 to output two-dimensional secondarily suppressed samples and the corresponding two-dimensional secondarily suppressed sample label set Set the element values in the two-dimensional secondarily suppressed sample label set to 1;

[0056] S5.6: Use the one-dimensional secondarily suppressed samples to train the one-dimensional quadratic generation network SG1, set the third generation loss function, optimize the one-dimensional quadratic generation network SG1, and obtain the first network parameter θ″ after the second update G1 and the optimized one-dimensional quadratic generation network SG′1;

[0057] Using the two-dimensional quadratic suppression samples Train the two-dimensional quadratic generation network SG2, set the fourth generation loss function, optimize the two-dimensional quadratic generation network SG2, and obtain the second network parameter θ″ after the second update G2 and the optimized two-dimensional quadratic generation network SG′2;

[0058] S5.7: Obtain the second optimized high sea state sea clutter generation neural network model by using the optimized one-dimensional quadratic generation network SG′1, the optimized one-dimensional quadratic discriminant network SD′1, the optimized two-dimensional quadratic generation network SG′2, and the optimized two-dimensional quadratic discriminant network SD′2.

[0059] Preferably, the specific method for training the generation network in steps S3.7 and S5.6 is as follows:

[0060] Using the one-dimensional quadratic generation sample set The specific method for training the one-dimensional generation network G1 is:

[0061] Input the one-dimensional quadratic generation sample set into the optimized one-dimensional discriminant network D′1 to obtain the first generation prediction label Using the first generation prediction label and the one-dimensional quadratic generation sample label set Calculate the first generation loss function value, optimize the one-dimensional generation network G1, and obtain the updated first network parameter θ′ G1 and the optimized one-dimensional generation network G′1;

[0062] Using the two-dimensional quadratic generation sample set The specific method for training the two-dimensional generation network G2 is:

[0063] Input the two-dimensional quadratic generation sample set into the optimized two-dimensional discriminant network D′2 to obtain the second generation prediction label Using the second generation prediction label and the two-dimensional quadratic generation sample label set Calculate the second generation loss function value, optimize the two-dimensional generation network G2, and obtain the updated second network parameter θ′ G2 and the optimized two-dimensional generation network G′2;

[0064] Using the one-dimensional quadratic suppression samples The specific method for training the one-dimensional quadratic generation network SG1 is:

[0065] Input the one-dimensional quadratic suppression samples Input the optimized one-dimensional quadratic discriminant network SD′1 to obtain the third generated prediction label Utilize the third generated prediction label and the one-dimensional quadratic suppression sample label set Calculate the value of the third loss function, optimize the one-dimensional quadratic generation network SG1, and obtain the first network parameter θ″ after quadratic update G1 and the optimized one-dimensional quadratic generation network SG′1;

[0066] The specific method for training the two-dimensional quadratic generation network SG2 using the two-dimensional quadratic suppression sample is as follows:

[0067] Input the two-dimensional quadratic suppression sample into the optimized two-dimensional quadratic discriminant network SD′2 to obtain the fourth generated prediction label Utilize the fourth generated prediction label and the two-dimensional quadratic suppression sample label set Calculate the value of the fourth generated loss function, optimize the two-dimensional quadratic generation network SG2, and obtain the second network parameter θ″ after quadratic update G2 and the optimized two-dimensional quadratic generation network SG′2.

[0068] Preferably, the loss functions in steps S3.5 and S3.7 are specifically:

[0069] The first discriminant loss function is specifically:

[0070]

[0071] where represents the value of the first discriminant loss function, x real and x noise respectively represent the data in the one-dimensional real sample set X real and the one-dimensional noise sample set X noise , and respectively represent the mathematical expectations of the one-dimensional real sample data and the one-dimensional noise sample data distribution functions;

[0072] The second discriminant loss function is specifically:

[0073]

[0074] where represents the value of the second discriminant loss function, z real and z noise respectively represent the data in the two-dimensional real sample set Z real and the two-dimensional noise sample set Z noise , and respectively represent the mathematical expectations of the two-dimensional real sample data and the two-dimensional noise sample data distribution functions;

[0075] The specific form of the first generation loss function is:

[0076]

[0077] wherein, represents the value of the first generation loss function, represents the data in the one-dimensional noise sample set obtained by secondary sampling, represents the mathematical expectation of the one-dimensional noise sample data distribution function obtained by secondary sampling;

[0078] The specific form of the second generation loss function is:

[0079]

[0080] wherein, represents the value of the second generation loss function, represents the data in the two-dimensional noise sample set obtained by secondary sampling, represents the mathematical expectation of the two-dimensional noise sample data distribution function obtained by secondary sampling.

[0081] Preferably, the loss functions in steps S5.4 and S5.6 are specifically:

[0082] The specific form of the third discriminant loss function is:

[0083]

[0084] wherein, is the value of the third discriminant loss function, sx real is the data in the measured low sea state sea clutter data set SX real in, d A1 is the data in the one-dimensional high sea state sea clutter data set D A1 in, and respectively represent the mathematical expectations of the measured low sea state sea clutter data and the one-dimensional high sea state sea clutter data distribution functions;

[0085] The specific form of the fourth discriminant loss function is:

[0086]

[0087] wherein, is the value of the fourth discriminant loss function, sz real is the two-dimensional low sea state sea clutter data in SZ real in, d A2 is DA2 The two-dimensional high sea state sea clutter data in and respectively represent the mathematical expectations of the distribution functions of two-dimensional low sea state sea clutter data and two-dimensional high sea state sea clutter data;

[0088] The specific form of the third generation loss function is:

[0089]

[0090] where is the value of the third generation loss function, represents the one-dimensional high sea state sea clutter data obtained by sub-sampling in the data set and represents the mathematical expectation of the distribution function of the one-dimensional high sea state sea clutter data obtained by sub-sampling;

[0091] The specific form of the fourth generation loss function is:

[0092]

[0093] where is the value of the fourth generation loss function, represents the two-dimensional high sea state sea clutter data obtained by sub-sampling in the data set and represents the mathematical expectation of the distribution function of the two-dimensional high sea state sea clutter data obtained by sub-sampling.

[0094] Preferably, in step S6, the high sea state sea clutter data to be suppressed is obtained, and the high sea state sea clutter generation neural network model optimized twice is used to suppress the high sea state sea clutter data to be suppressed. The specific method is as follows:

[0095] S6.1: Measure and obtain the high sea state sea clutter data T0 to be suppressed;

[0096] S6.2: Input the high sea state sea clutter data T0 to be suppressed into the optimized one-dimensional secondary generation network SG′1 to obtain the one-dimensional suppressed high sea state sea clutter data T1;

[0097] S6.3: Perform time-frequency spectrum conversion on the one-dimensional suppressed high sea state sea clutter data T1 to obtain two-dimensional high sea state sea clutter data T2;

[0098] S6.4: Input the two-dimensional high sea state sea clutter data T2 into the optimized two-dimensional secondary generation network SG′2 for two-dimensional suppression to obtain the suppressed high sea state sea clutter data.

[0099] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0100] The present invention provides a method for suppressing high sea state sea clutter based on deep learning, including: obtaining a measured sea clutter data set; obtaining a simulated sea clutter data set and a measured high sea state sea clutter data set according to the measured sea clutter data set; inputting the simulated sea clutter data set and the measured high sea state sea clutter data set into a preset high sea state sea clutter generation neural network model for adversarial training to obtain a first optimized high sea state sea clutter generation neural network model; using the first optimized high sea state sea clutter generation neural network model to obtain a one-dimensional high sea state sea clutter data set and a two-dimensional high sea state sea clutter data set; inputting the one-dimensional high sea state sea clutter data set and the two-dimensional high sea state sea clutter data set into the first optimized high sea state sea clutter generation neural network model for adversarial training to obtain a second optimized high sea state sea clutter generation neural network model; obtaining high sea state sea clutter data to be suppressed, and using the second optimized high sea state sea clutter generation neural network model to suppress the high sea state sea clutter data to be suppressed;

[0101] This method takes into account both measured and simulated sea clutter data. On the one hand, it reduces the demand for measured high sea state sea clutter samples and reduces the time-consuming and laborious process of obtaining measured data. On the other hand, it helps the model comprehensively learn the physical characteristics and statistical distribution characteristics of high sea state sea clutter, generates more realistic high sea state sea clutter data samples, and establishes a sea clutter model with stronger generalization ability;

[0102] In addition, this method suppresses high sea state sea clutter data based on deep learning and simultaneously retrains the high sea state sea clutter generation model, which can provide a large number of samples that fit the real high sea state sea clutter data distribution for the high sea state sea clutter suppression model and reduce the overfitting problem in the learning process of the suppression model;

[0103] In addition, this method combines the one-dimensional characteristics and two-dimensional characteristics of sea clutter, suppresses sea clutter in two dimensions, learns more abundant features, improves the detection efficiency of targets under complex sea conditions, and significantly improves the suppression effect of high sea state sea clutter. Description of the Drawings

[0104] Figure 1 It is a flowchart of a method for suppressing high sea state sea clutter based on deep learning provided in Embodiment 1.

[0105] Figure 2 It is a structural diagram of the high sea state sea clutter generation network G provided in Embodiment 2.

[0106] Figure 3 It is a structural diagram of the high sea state sea clutter discrimination network D provided in Embodiment 2. Detailed Embodiment

[0107] The drawings are only for illustrative purposes and should not be construed as a limitation of this patent;

[0108] To better illustrate this embodiment, some components in the drawings are omitted, enlarged or reduced, which does not represent the dimensions of the actual product;

[0109] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0110] The technical solution of the present invention will be further described below with reference to the drawings and embodiments.

[0111] Embodiment 1

[0112] As Figure 1 shown, this embodiment provides a high-sea-state sea clutter suppression method based on deep learning, including the following steps:

[0113] S1: Obtain the measured sea clutter data set;

[0114] S2: According to the measured sea clutter data set, obtain the simulated sea clutter data set and the measured high-sea-state sea clutter data set;

[0115] S3: Input the simulated sea clutter data set and the measured high-sea-state sea clutter data set into a preset high-sea-state sea clutter generation neural network model for adversarial training to obtain a first-optimized high-sea-state sea clutter generation neural network model;

[0116] S4: Use the first-optimized high-sea-state sea clutter generation neural network model to obtain a one-dimensional high-sea-state sea clutter data set and a two-dimensional high-sea-state sea clutter data set;

[0117] S5: Input the one-dimensional high-sea-state sea clutter data set and the two-dimensional high-sea-state sea clutter data set into the first-optimized high-sea-state sea clutter generation neural network model for adversarial training to obtain a second-optimized high-sea-state sea clutter generation neural network model;

[0118] S6: Obtain the high-sea-state sea clutter data to be suppressed, and use the second-optimized high-sea-state sea clutter generation neural network model to suppress the high-sea-state sea clutter data to be suppressed.

[0119] In the specific implementation process, first, an actual measured sea clutter data set is obtained through actual measurement. Then, according to the actual measured sea clutter data set, a simulated sea clutter data set and an actual measured high-sea-state sea clutter data set are obtained. The simulated sea clutter data set and the actual measured high-sea-state sea clutter data set are input into a preset high-sea-state sea clutter generation neural network model for adversarial training to obtain a first-optimized high-sea-state sea clutter generation neural network model. The first-optimized high-sea-state sea clutter generation neural network model is used to obtain a one-dimensional high-sea-state sea clutter data set and a two-dimensional high-sea-state sea clutter data set. The one-dimensional high-sea-state sea clutter data set and the two-dimensional high-sea-state sea clutter data set are input into the first-optimized high-sea-state sea clutter generation neural network model for adversarial training to obtain a second-optimized high-sea-state sea clutter generation neural network model. Finally, the high-sea-state sea clutter data to be suppressed is obtained, and the second-optimized high-sea-state sea clutter generation neural network model is used to suppress the high-sea-state sea clutter data to be suppressed.

[0120] This method takes into account both actual measured and simulated sea clutter data. On the one hand, it reduces the need for actual measured high-sea-state sea clutter samples and reduces the time-consuming and laborious process of obtaining actual measured data. On the other hand, it helps the model comprehensively learn the physical characteristics and statistical distribution characteristics of high-sea-state sea clutter, generate more realistic high-sea-state sea clutter data samples, and establish a sea clutter model with stronger generalization ability. In addition, this method suppresses high-sea-state sea clutter data based on deep learning and simultaneously retrains the high-sea-state sea clutter generation model, which can provide a large number of samples that fit the real high-sea-state sea clutter data distribution for the high-sea-state sea clutter suppression model and reduce the overfitting problem in the learning process of the suppression model.

[0121] Embodiment 2

[0122] This embodiment provides a method for suppressing high-sea-state sea clutter based on deep learning, including the following steps:

[0123] S1: Obtain an actual measured sea clutter data set;

[0124] S2: According to the actual measured sea clutter data set, obtain a simulated sea clutter data set and an actual measured high-sea-state sea clutter data set:

[0125] S3: Input the simulated sea clutter data set and the actual measured high-sea-state sea clutter data set into a preset high-sea-state sea clutter generation neural network model for adversarial training to obtain a first-optimized high-sea-state sea clutter generation neural network model;

[0126] S4: Use the first-optimized high-sea-state sea clutter generation neural network model to obtain a one-dimensional high-sea-state sea clutter data set and a two-dimensional high-sea-state sea clutter data set;

[0127] S5: Input the one-dimensional high sea state sea clutter dataset and the two-dimensional high sea state sea clutter dataset into the once-optimized high sea state sea clutter generation neural network model for adversarial training to obtain a twice-optimized high sea state sea clutter generation neural network model;

[0128] S6: Obtain the high sea state sea clutter data to be suppressed, and use the twice-optimized high sea state sea clutter generation neural network model to suppress the high sea state sea clutter data to be suppressed.

[0129] In the specific implementation process, first obtain 500 pieces of measured sea clutter data through actual measurement and integrate them to obtain a measured sea clutter dataset D = {d1, d2, ···, d 500};

[0130] Divide the 500 pieces of measured sea clutter data according to the sea state level a at the time of data collection, and set a sea state level threshold A. If a < A, then divide the measured sea clutter data into the measured low sea state sea clutter dataset D L , otherwise, divide the measured sea clutter data into the measured high sea state sea clutter dataset D H ; In this embodiment, A = 5, and finally 400 pieces of low sea state sea clutter data and 100 pieces of high sea state sea clutter datasets

[0131] After that, randomly sample 100 low sea state sea clutter data samples from the measured low sea state sea clutter dataset D L , and respectively superimpose 3 kinds of simulated sea clutter based on the K distribution, lognormal distribution, and Weibull distribution to obtain 300 pieces of simulated sea clutter data

[0132] This embodiment presupposes a high sea state sea clutter generation neural network model, which specifically includes a high sea state sea clutter generation network G and a high sea state sea clutter discrimination network D;

[0133] As Figure 2 shown, the high sea state sea clutter generation network G includes a dimension initialization module, a sea clutter encoding module, several deformable attention conversion modules, and a sea clutter decoding module connected in sequence;

[0134] The dimension initialization module includes a first fully connected layer and a first Reshape layer connected in sequence;

[0135] The sea clutter encoding module includes a first convolutional layer, a first batch normalization layer, a first activation layer, and a downsampling layer connected in sequence;

[0136] Each of the deformable attention conversion modules includes a first deformable convolutional layer, a second batch normalization layer, a second activation layer, an attention layer, a weighted product point, and a first residual addition point connected in sequence; the output end of the second activation layer is also connected to the input end of the weighted product point, and the input end of the first deformable convolutional layer is also connected to the input end of the first residual addition point;

[0137] The sea clutter decoding module includes an upsampling layer, a third batch normalization layer, a third activation layer, a second convolutional layer, and a fourth activation layer connected in sequence.

[0138] As Figure 3 shown, the high sea state sea clutter discrimination network D includes a plurality of sea clutter downsampling modules, a plurality of spectral normalization convolutional modules, a plurality of deformable convolutional modules, and a sea clutter classification module connected in sequence;

[0139] Each of the sea clutter downsampling modules includes a third convolutional layer, a fourth batch normalization layer, a fifth activation layer, and a second residual addition point connected in sequence; and the input end of the third convolutional layer is also connected to the input end of the second residual addition point;

[0140] The spectral normalization convolutional module includes a fourth convolutional layer, a spectral normalization layer, and a sixth activation layer connected in sequence;

[0141] The deformable convolutional module includes a second deformable convolutional layer and a seventh activation layer connected in sequence;

[0142] The sea clutter classification module includes a second Reshape layer, a second fully connected layer, and an eighth activation layer connected in sequence;

[0143] Input the simulated sea clutter dataset D S and the measured high sea state sea clutter dataset D H into a preset high sea state sea clutter generation neural network model for adversarial training to obtain a once-optimized high sea state sea clutter generation neural network model. The specific method is as follows:

[0144] S3.1: Use the union of the simulated sea clutter dataset D S and the measured high sea state sea clutter dataset D H as a one-dimensional real sample set Perform time-frequency conversion on the one-dimensional real sample set X real to obtain a two-dimensional real sample set where the calculation formula of the short-time Fourier transform is:

[0145]

[0146] where ω(t) is the window function, x(t) is the signal to be transformed, and X(t,ω) is the Fourier transform of ω(t - τ)x(τ);

[0147] S3.2: Construct a one-dimensional noise sample set X that satisfies the Gaussian distribution noise and a two-dimensional noise sample set Z that satisfies the Gaussian distribution noise , where and and

[0148] S3.3: Set the first network parameter θ G1 and the second network parameter θ G2 and assign them to the high sea state sea clutter generation network G to obtain a one-dimensional generation network G1 and a two-dimensional generation network G2. Set the third network parameter θ D1 and the fourth network parameter θ D2 and assign them to the high sea state sea clutter discrimination network D to obtain a one-dimensional discrimination network D1 and a two-dimensional discrimination network D2, where, θ G1 , θ D1 , θ G2 , In this embodiment, the number of sampling samples b for a single training is 50;

[0149] S3.4: Randomly sample the one-dimensional noise sample set X noise to obtain a one-dimensional sampled noise sample set and input it into the one-dimensional generation network G1 to obtain a one-dimensional generated sample set and the corresponding one-dimensional generated sample label set Set the element values in the one-dimensional generated sample label set to 0; Randomly sample the one-dimensional real sample set X real to obtain a one-dimensional sampled real sample set and the corresponding one-dimensional real sample label set Set the element values in the one-dimensional real sample label set to 1; Randomly sample the two-dimensional noise sample set Z noise and to obtain a two-dimensional sampled noise sample set and input it into the two-dimensional generation network G2 to obtain a two-dimensional generated sample set and the corresponding two-dimensional generated sample label set Set the element values in the two-dimensional generated sample label set to 0; Randomly sample the two-dimensional real sample set Z real to obtain a two-dimensional sampled real sample set and the corresponding two-dimensional real sample label set Set the element values in the two-dimensional real sample label set to 1;

[0150] S3.5: Utilize the one-dimensional generated sample set and the one-dimensional sampled real sample set to train the one-dimensional discriminant network D1, set the first discriminant loss function, optimize the one-dimensional discriminant network D1, and obtain the updated third network parameter θ′ D1 and the optimized one-dimensional discriminant network D′1, specifically:

[0151] Input the one-dimensional generated sample set and the one-dimensional sampled real sample set into the one-dimensional discriminant network D1, and output the first discriminant prediction label set Input the one-dimensional generated sample label set the one-dimensional real sample label set and the first discriminant prediction label set into the first discriminant loss function Calculate the first discriminant loss, optimize the one-dimensional discriminant network D1, and obtain the updated third network parameter θ′ D1 and the optimized one-dimensional discriminant network D′1, where the first discriminant loss function is specifically:

[0152]

[0153] where represents the value of the first discriminant loss function, x real and x noise respectively represent the data in the one-dimensional real sample set X real and the one-dimensional noise sample set X noise ; and respectively represent the mathematical expectations of the one-dimensional real sample data and the one-dimensional noise sample data distribution functions;

[0154] Utilize the two-dimensional generated sample set and the two-dimensional sampled real sample set to train the two-dimensional discriminant network D2, set the second discriminant loss function, optimize the two-dimensional discriminant network D2, and obtain the updated fourth network parameter θ′ D2 and the optimized two-dimensional discriminant network D′2, specifically:

[0155] Input the two-dimensional generated sample set and the two-dimensional sampled real sample set into the two-dimensional discriminant network D2, and output the second discriminant prediction label set Input the two-dimensional generated sample label set the two-dimensional real sample label set and the second discriminant prediction label set Input the second discriminant loss function Calculate the second discriminant loss, optimize the two-dimensional discriminant network D2, and obtain the updated fourth network parameter θ′ D2 And the optimized two-dimensional discriminant network D′2, where the second discriminant loss function is specifically:

[0156]

[0157] Wherein, Represents the value of the second discriminant loss function, z real And z noise Respectively represent the two-dimensional real sample set Z real And the two-dimensional noise sample set Z noise In the data, And Respectively represent the mathematical expectations of the two-dimensional real sample data and the two-dimensional noise sample data distribution functions;

[0158] S3.6: Randomly sample the one-dimensional noise sample set X noise Again to obtain a one-dimensional secondary sampled noise sample set Input the one-dimensional secondary sampled noise sample set Into the one-dimensional generation network G1, and output a one-dimensional secondary generated sample set And the corresponding one-dimensional secondary generated sample label set Set the element values in the one-dimensional secondary generated sample label set To 1;

[0159] Randomly sample the two-dimensional noise sample set Z noise Again to obtain a two-dimensional secondary sampled noise sample set Input the two-dimensional secondary sampled noise sample set Into the two-dimensional generation network G2, and output a two-dimensional secondary generated sample set And the corresponding two-dimensional secondary generated sample label set Set the element values in the two-dimensional secondary generated sample label set To 1;

[0160] S3.7: Use the one-dimensional secondary generated sample set Train the one-dimensional generation network G1, and set the first generation loss function to optimize the one-dimensional generation network G1 to obtain the updated first network parameter θ′ G1 And the optimized one-dimensional generation network G′1, specifically:

[0161] Input the one-dimensional secondary generated sample set Input the optimized one-dimensional discriminant network D′1 to obtain the first generated prediction label Utilize the first generated prediction label and the one-dimensional quadratic generated sample label set Calculate the value of the first generation loss function, optimize the one-dimensional generation network G1, and obtain the updated first network parameter θ′ G1 and the optimized one-dimensional generation network G′1, where the first generation loss function is specifically:

[0162]

[0163] where represents the value of the first generation loss function, represents the data in the one-dimensional noise sample set of quadratic sampling, represents the mathematical expectation of the one-dimensional noise sample data distribution function of quadratic sampling;

[0164] Utilize the two-dimensional quadratic generated sample set Train the two-dimensional generation network G2, set the second generation loss function, optimize the two-dimensional generation network G2, and obtain the updated second network parameter θ′ G2 and the optimized two-dimensional generation network G′2, specifically:

[0165] Input the two-dimensional quadratic generated sample set into the optimized two-dimensional discriminant network D′2 to obtain the second generated prediction label Utilize the second generated prediction label and the two-dimensional quadratic generated sample label set Calculate the value of the second generation loss function, optimize the two-dimensional generation network G2, and obtain the updated second network parameter θ′ G2 and the optimized two-dimensional generation network G′2, where the second generation loss function is specifically:

[0166]

[0167] where represents the value of the second generation loss function, represents the data in the two-dimensional noise sample set of quadratic sampling, represents the mathematical expectation of the two-dimensional noise sample data distribution function of quadratic sampling;

[0168] S3.8: Utilize the optimized one-dimensional generation network G′1, the optimized one-dimensional discriminant network D′1, the optimized two-dimensional generation network G′2, and the optimized two-dimensional discriminant network D′2 to obtain a once-optimized high sea state sea clutter generation neural network model.

[0169] After that, input the one-dimensional noise sample set Xnoise Input the optimized one - dimensional generation network \(G'_1\) to obtain 400 one - dimensional high sea - state sea clutter data

[0170] Input the two - dimensional noise sample set \(Z\) noise Input the optimized two - dimensional generation network \(G'_2\) to obtain 400 two - dimensional high sea - state sea clutter data

[0171] Then input the one - dimensional high sea - state sea clutter data set \(D\) A1 and the two - dimensional high sea - state sea clutter data set \(D\) A2 into the once - optimized high sea - state sea clutter generation neural network model for adversarial training to obtain the twice - optimized high sea - state sea clutter generation neural network model. Specifically:

[0172] S5.1: Perform time - frequency spectrum conversion on the measured low sea - state sea clutter data set \(D\) L to obtain the two - dimensional low sea - state sea clutter sample set \(SZ\) real ;

[0173] S5.2: Assign the updated first network parameter \(\theta'\) G1 and the updated second network parameter \(\theta'\) G2 to the high sea - state sea clutter generation network \(G\) to obtain the one - dimensional twice - generation network \(SG_1\) and the two - dimensional twice - generation network \(SG_2\). Assign the updated third network parameter \(\theta'\) D1 and the fourth network parameter \(\theta'\) D2 to the high sea - state sea clutter discrimination network \(D\) to obtain the one - dimensional twice - discrimination network \(SD_1\) and the two - dimensional twice - discrimination network \(SD_2\). In this embodiment, the number of sampling samples per single training \(c = 50\);

[0174] S5.3: Randomly sample the one - dimensional high sea - state sea clutter data set \(D\) A1 to obtain one - dimensional sea clutter samples Input the one - dimensional sea clutter samples into the one - dimensional twice - generation network \(SG_1\) to obtain one - dimensional suppression samples and the corresponding one - dimensional suppression sample label set Set the element values in the one - dimensional suppression sample label set to 0; Randomly sample the measured low sea - state sea clutter data set \(D\) L to obtain one - dimensional low sea - state samples and the corresponding one - dimensional low sea - state sample label set Set the element values in the one - dimensional low sea - state sample label set to 1;

[0175] Randomly sample the two - dimensional high sea - state sea clutter data set \(D\)A2 Perform random sampling to obtain two-dimensional sea clutter samples Input the two-dimensional sea clutter samples into the two-dimensional secondary generation network SG2 to obtain two-dimensional suppression samples and the corresponding two-dimensional suppression sample label set Set the element values in the two-dimensional suppression sample label set to 0; For the two-dimensional low sea state sea clutter sample set SZ real Perform random sampling to obtain two-dimensional low sea state samples and the corresponding two-dimensional low sea state sample label set Set the element values in the two-dimensional low sea state sample label set to 1;

[0176] S5.4: Use the one-dimensional suppression samples and one-dimensional low sea state samples to train the one-dimensional secondary discriminant network SD1, and set the third discriminant loss function to optimize the one-dimensional secondary discriminant network SD1 to obtain the third network parameters θ″ after secondary update D1 and the optimized one-dimensional secondary discriminant network SD′1, specifically:

[0177] Input the one-dimensional suppression samples and one-dimensional low sea state samples into the one-dimensional secondary discriminant network SD1, and output the third discriminant prediction label Input the one-dimensional suppression sample label set the one-dimensional low sea state sample label set and the third discriminant prediction label into the third discriminant loss function Calculate the third discriminant loss, optimize the one-dimensional secondary discriminant network SD1, and obtain the third network parameters θ″ after secondary update D1 and the optimized one-dimensional secondary discriminant network SD′1, where the third discriminant loss function is specifically:

[0178]

[0179] where, is the value of the third discriminant loss function, sx real is the data in the measured low sea state sea clutter data set SX real in, d A1 is the data in the one-dimensional high sea state sea clutter data set D A1 in, and respectively represent the mathematical expectations of the measured low sea state sea clutter data and the one-dimensional high sea state sea clutter data distribution functions;

[0180] Utilize the two-dimensional suppression samples and two-dimensional low sea state samples Train the two-dimensional quadratic discriminant network SD2, set the fourth discriminant loss function, optimize the two-dimensional quadratic discriminant network SD2, and obtain the fourth network parameters θ″ after the second update D2 and the optimized two-dimensional quadratic discriminant network SD′2, specifically:

[0181] Input the two-dimensional suppression samples and two-dimensional low sea state samples into the two-dimensional quadratic discriminant network SD2, and output the fourth discriminant prediction label set Input the two-dimensional suppression sample label set the two-dimensional low sea state sample label set and the fourth discriminant prediction label set into the fourth discriminant loss function Calculate the fourth discriminant loss, optimize the two-dimensional quadratic discriminant network SD2, and obtain the fourth network parameters θ″ after the second update D2 and the optimized two-dimensional quadratic discriminant network SD′2, where the fourth discriminant loss function is specifically:

[0182]

[0183] where, is the value of the fourth discriminant loss function, sz real is the two-dimensional low sea state sea clutter data in SZ real and d A2 is the two-dimensional high sea state sea clutter data in D A2 , and respectively represent the mathematical expectations of the two-dimensional low sea state sea clutter data and the two-dimensional high sea state sea clutter data distribution functions;

[0184] S5.5: Randomly sample the one-dimensional high sea state sea clutter dataset D A1 again to obtain one-dimensional secondary sampled sea clutter samples Input the one-dimensional secondary sampled sea clutter samples into the one-dimensional secondary generation network SG1 to obtain one-dimensional secondary suppression samples and the corresponding one-dimensional secondary suppression sample label set Set the element values in the one-dimensional secondary suppression sample label set to 1;

[0185] Randomly sample the two-dimensional high sea state sea clutter dataset D A2 again to obtain two-dimensional secondary sampled sea clutter samples Input the two-dimensional quadratic sampled sea clutter samples into the two-dimensional quadratic generation network SG2, and output two-dimensional quadratic suppression samples and the corresponding two-dimensional quadratic suppression sample label set Set the element values in the two-dimensional quadratic suppression sample label set to 1;

[0186] S5.6: Use the one-dimensional quadratic suppression samples to train the one-dimensional quadratic generation network SG1, and set the third generation loss function to optimize the one-dimensional quadratic generation network SG1 to obtain the first network parameter θ″ after secondary update G1 and the optimized one-dimensional quadratic generation network SG′1, specifically:

[0187] Input the one-dimensional quadratic suppression samples into the optimized one-dimensional quadratic discriminant network SD′1 to obtain the third generation prediction label Use the third generation prediction label and the one-dimensional quadratic suppression sample label set to calculate the value of the third loss function, optimize the one-dimensional quadratic generation network SG1, and obtain the first network parameter θ″ after secondary update G1 and the optimized one-dimensional quadratic generation network SG′1, where the third generation loss function is specifically:

[0188]

[0189] where, is the value of the third generation loss function, represents the dataset in the one-dimensional high sea state sea clutter data of quadratic sampling, represents the mathematical expectation of the one-dimensional high sea state sea clutter data distribution function of quadratic sampling;

[0190] Use the two-dimensional quadratic suppression samples to train the two-dimensional quadratic generation network SG2, and set the fourth generation loss function to optimize the two-dimensional quadratic generation network SG2 to obtain the second network parameter θ″ after secondary update G2 and the optimized two-dimensional quadratic generation network SG′2, specifically:

[0191] Input the two-dimensional quadratic suppression samples into the optimized two-dimensional quadratic discriminant network SD′2 to obtain the fourth generation prediction label Use the fourth generation prediction label and the two-dimensional quadratic suppression sample label set Calculate the value of the fourth generation loss function, optimize the two-dimensional quadratic generation network SG2, and obtain the second network parameter θ″ after the second update G2 and the optimized two-dimensional quadratic generation network SG′2, where the fourth generation loss function is specifically:

[0192]

[0193] where, is the value of the fourth generation loss function, represents the dataset in the two-dimensional high sea state sea clutter data of the second sampling, represents the mathematical expectation of the distribution function of the two-dimensional high sea state sea clutter data of the second sampling;

[0194] S5.7: Obtain the second optimized high sea state sea clutter generation neural network model by using the optimized one-dimensional quadratic generation network SG′1, the optimized one-dimensional quadratic discriminant network SD′1, the optimized two-dimensional quadratic generation network SG′2, and the optimized two-dimensional quadratic discriminant network SD′2;

[0195] Finally, obtain the high sea state sea clutter data to be suppressed, and use the second optimized high sea state sea clutter generation neural network model to suppress the high sea state sea clutter data to be suppressed. The specific method is:

[0196] S6.1: Measure and obtain the high sea state sea clutter data T0 to be suppressed;

[0197] S6.2: Input the high sea state sea clutter data T0 to be suppressed into the optimized one-dimensional quadratic generation network SG′1 to obtain the one-dimensional suppressed high sea state sea clutter data T1;

[0198] S6.3: Perform time-frequency spectrum conversion on the one-dimensional suppressed high sea state sea clutter data T1 to obtain the two-dimensional high sea state sea clutter data T2;

[0199] S6.4: Input the two-dimensional high sea state sea clutter data T2 into the optimized two-dimensional quadratic generation network SG′2 for two-dimensional suppression to obtain the suppressed high sea state sea clutter data;

[0200] This method simultaneously considers measured and simulated sea clutter data. On the one hand, it reduces the demand for measured high-sea-state sea clutter samples and the time-consuming and laborious process of obtaining measured data. On the other hand, it helps the model comprehensively learn the physical and statistical distribution characteristics of high-sea-state sea clutter, generate more realistic high-sea-state sea clutter data samples, and establish a sea clutter model with stronger generalization ability. In addition, this method suppresses high-sea-state sea clutter data based on deep learning and retrains the high-sea-state sea clutter generation model, which can provide a large number of samples that fit the real high-sea-state sea clutter data distribution for the high-sea-state sea clutter suppression model and reduce the overfitting problem in the learning process of the suppression model. Moreover, this method combines the one-dimensional and two-dimensional characteristics of sea clutter to suppress sea clutter in two dimensions, learn more abundant features, improve the detection efficiency of targets under complex sea conditions, and significantly improve the suppression effect of high-sea-state sea clutter.

[0201] Like reference numerals correspond to like components;

[0202] The terms used to describe the positional relationship in the drawings are for illustrative purposes only and should not be construed as a limitation of this patent;

[0203] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention and are not intended to limit the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for suppressing sea clutter in high sea states based on deep learning, characterized in that, It includes the following steps: S1: Obtain the measured sea clutter data set; S2: According to the measured sea clutter data set, obtain the simulated sea clutter data set and the measured high-sea-state sea clutter data set; S3: Input the simulated sea clutter data set and the measured high-sea-state sea clutter data set into a preset high-sea-state sea clutter generation neural network model for adversarial training to obtain a first-optimized high-sea-state sea clutter generation neural network model; S4: Use the first-optimized high-sea-state sea clutter generation neural network model to obtain a one-dimensional high-sea-state sea clutter data set and a two-dimensional high-sea-state sea clutter data set; S5: Input the one-dimensional high-sea-state sea clutter data set and the two-dimensional high-sea-state sea clutter data set into the first-optimized high-sea-state sea clutter generation neural network model for adversarial training to obtain a second-optimized high-sea-state sea clutter generation neural network model; S6: Obtain the high-sea-state sea clutter data to be suppressed, and use the second-optimized high-sea-state sea clutter generation neural network model to suppress the high-sea-state sea clutter data to be suppressed.

2. The high sea clutter suppression method based on deep learning according to claim 1, characterized in that In step S2, the specific method for obtaining the simulated sea clutter data set and the measured high-sea-state sea clutter data set according to the measured sea clutter data set is as follows: S2.1: Divide the measured sea clutter data set D into a measured low sea state sea clutter data set and a measured high sea state sea clutter data set ; S2.2: Superimpose simulated sea clutter with several distributions on the measured sea clutter dataset under low sea state conditions to obtain a simulated sea clutter dataset .

3. The high sea state sea clutter suppression method based on deep learning according to claim 2, wherein, In step S3, the preset high-sea-state sea clutter generation neural network model is specifically: The preset high-sea-state sea clutter generation neural network model includes a high-sea-state sea clutter generation network G and a high-sea-state sea clutter discrimination network D; The high-sea-state sea clutter generation network G includes a dimension initialization module, a sea clutter encoding module, several deformable attention conversion modules, and a sea clutter decoding module connected in sequence; The dimension initialization module includes a first fully connected layer and a first Reshape layer connected in sequence; The sea clutter encoding module includes a first convolutional layer, a first batch normalization layer, a first activation layer, and a downsampling layer connected in sequence; Each of the deformable attention conversion modules includes a first deformable convolutional layer, a second batch normalization layer, a second activation layer, an attention layer, a weighted product point, and a first residual addition point connected in sequence; the output end of the second activation layer is also connected to the input end of the weighted product point, and the input end of the first deformable convolutional layer is also connected to the input end of the first residual addition point; The sea clutter decoding module includes an upsampling layer, a third batch normalization layer, a third activation layer, a second convolutional layer, and a fourth activation layer connected in sequence; The high-sea-state sea clutter discrimination network D includes several sea clutter downsampling modules, several spectral normalization convolutional modules, several deformable convolutional modules, and a sea clutter classification module connected in sequence; Each of the sea clutter downsampling modules includes a third convolutional layer, a fourth batch normalization layer, a fifth activation layer, and a second residual addition point connected in sequence; and the input end of the third convolutional layer is also connected to the input end of the second residual addition point; The spectral normalization convolutional module includes a fourth convolutional layer, a spectral normalization layer, and a sixth activation layer connected in sequence; The deformable convolutional module includes a second deformable convolutional layer and a seventh activation layer connected in sequence; The sea clutter classification module includes a second Reshape layer, a second fully connected layer, and an eighth activation layer connected in sequence.

4. A method for suppressing sea clutter in high sea states based on deep learning according to claim 3, characterized in that, In step S3, the simulated sea clutter data set and the measured high sea state sea clutter data set are input into a preset high sea state sea clutter generation neural network model for adversarial training to obtain a once-optimized high sea state sea clutter generation neural network model. The specific method is as follows: S3.1: Use the union of the simulated sea clutter data set and the measured high sea state sea clutter data set as the one-dimensional true sample set , for the one-dimensional true sample set Perform time-frequency spectrum conversion to obtain a two-dimensional true sample set ; S3.2: Construct a one-dimensional noise sample set that satisfies the Gaussian distribution and a two-dimensional noise sample set that satisfies the Gaussian distribution respectively and ; S3.3: Set the first network parameters respectively and assign them to the high sea state sea clutter generation network G to obtain a one-dimensional generation network G1 and a two-dimensional generation network G2. Set the third network parameters respectively and assign them to the high sea state sea clutter discrimination network D to obtain a one-dimensional discrimination network D1 and a two-dimensional discrimination network D2; S3.4: For the one-dimensional noise sample set perform random sampling to obtain a one-dimensional sampled noise sample set , and input it into the one-dimensional generation network G1 to obtain a one-dimensional generated sample set and the corresponding one-dimensional generated sample label set . Set the element values in the one-dimensional generated sample label set to 0. Perform random sampling on the one-dimensional real sample set to obtain a one-dimensional sampled real sample set and the corresponding one-dimensional real sample label set . Set the element values in the one-dimensional real sample label set to 1; Perform random sampling on the two-dimensional noise sample set to obtain a two-dimensional sampled noise sample set , and input it into the two-dimensional generation network G2 to obtain a two-dimensional generated sample set and the corresponding two-dimensional generated sample label set . Set the element values in the two-dimensional generated sample label set to 0. Perform random sampling on the two-dimensional real sample set to obtain a two-dimensional sampled real sample set and the corresponding two-dimensional real sample label set . Set the element values in the two-dimensional real sample label set to 1; where b is the number of sampled samples; S3.5: Utilize the one-dimensional generated sample set and the one-dimensional sampled real sample set to train the one-dimensional discriminant network D1, set the first discriminant loss function, optimize the one-dimensional discriminant network D1, and obtain the updated third network parameters and the optimized one-dimensional discriminant network ; Using the two-dimensional generated sample set and the two-dimensional sampled real sample set Train the two-dimensional discriminant network D2, set the second discriminant loss function, optimize the two-dimensional discriminant network D2, and obtain the updated fourth network parameters and the optimized two-dimensional discriminant network ; S3.6: For the one-dimensional noise sample set perform random sampling again to obtain a one-dimensional secondary sampling noise sample set . Input the one-dimensional secondary sampling noise sample set into the one-dimensional generation network G1 to output a one-dimensional secondary generation sample set and the corresponding one-dimensional secondary generation sample label set . Set the element values in the one-dimensional secondary generation sample label set to 1. For the two-dimensional noise sample set Perform random sampling again to obtain a two-dimensional secondary sampling noise sample set , and input the two-dimensional secondary sampling noise sample set into the two-dimensional generation network G2 to output a two-dimensional secondary generation sample set and the corresponding two-dimensional secondary generation sample label set . Set the element values in the two-dimensional secondary generation sample label set to 1; S3.7: Utilize the one-dimensional quadratic generated sample set Train the one-dimensional generation network G1, set the first generation loss function, optimize the one-dimensional generation network G1, and obtain the updated first network parameters and the optimized one-dimensional generation network ; Utilize the two-dimensional quadratic generated sample set Train the two-dimensional generation network G2, set the second generation loss function, optimize the two-dimensional generation network G2, and obtain the updated second network parameters and the optimized two-dimensional generation network ; S3.8: Utilize the optimized one-dimensional generation network , the optimized one-dimensional discriminative network , the optimized two-dimensional generation network and the optimized two-dimensional discriminative network to obtain a once-optimized high sea state sea clutter generation neural network model.

5. A method for suppressing sea clutter in high sea states based on deep learning according to claim 4, characterized in that, In step S4, using the once-optimized high sea state sea clutter generation neural network model, a one-dimensional high sea state sea clutter data set and a two-dimensional high sea state sea clutter data set are obtained. The specific method is as follows: Input the one-dimensional noise sample set into the optimized one-dimensional generation network to obtain a one-dimensional high sea state sea clutter data set ; Input the two-dimensional noise sample set into the optimized two-dimensional generation network to obtain a two-dimensional high sea state sea clutter data set .

6. A method for suppressing sea clutter in high sea states based on deep learning according to claim 5, characterized in that, In step S5, the one-dimensional high sea state sea clutter data set and the two-dimensional high sea state sea clutter data set are input into the once-optimized high sea state sea clutter generation neural network model for adversarial training to obtain a twice-optimized high sea state sea clutter generation neural network model. The specific method is as follows: S5.1: Perform time-frequency spectrum conversion on the measured low sea state sea clutter data set to obtain a two-dimensional low sea state sea clutter sample set ; S5.2: Respectively assign the updated first network parameters and the updated second network parameters to the high sea state sea clutter generation network G to obtain the one-dimensional quadratic generation network SG1 and the two-dimensional quadratic generation network SG2. Respectively assign the updated third network parameters to the high sea state sea clutter discrimination network D to obtain the one-dimensional quadratic discrimination network SD1 and the two-dimensional quadratic discrimination network SD2; S5.3: For the one-dimensional high sea state sea clutter data set perform random sampling to obtain one-dimensional sea clutter samples , and input the one-dimensional sea clutter samples into the one-dimensional quadratic generation network SG1 to obtain one-dimensional suppression samples and the corresponding one-dimensional suppression sample label set . Set the element values in the one-dimensional suppression sample label set to 0; For the measured low sea state sea clutter data set Perform random sampling to obtain one-dimensional low sea state samples And the corresponding one-dimensional low sea state sample label set , Set the element values in the one-dimensional low sea state sample label set To 1; For the two-dimensional high sea state sea clutter data set Perform random sampling to obtain two-dimensional sea clutter samples , and input the two-dimensional sea clutter samples into the two-dimensional quadratic generation network SG2 to obtain two-dimensional suppression samples and the corresponding two-dimensional suppression sample label set . Set the element values in the two-dimensional suppression sample label set to 0; For the two-dimensional low sea state sea clutter sample set Perform random sampling to obtain two-dimensional low sea state samples And the corresponding two-dimensional low sea state sample label set , set the element values in the two-dimensional low sea state sample label set To 1; where r is the number of sampled samples; S5.4: Utilize the one-dimensional suppression samples and one-dimensional low sea state samples to train the one-dimensional quadratic discriminant network SD1, set the third discriminant loss function, optimize the one-dimensional quadratic discriminant network SD1, and obtain the third network parameters after the second update and the optimized one-dimensional quadratic discriminant network ; Using the two-dimensional suppression samples and two-dimensional low sea state samples Train the two-dimensional quadratic discriminant network SD2, set the fourth discriminant loss function, optimize the two-dimensional quadratic discriminant network SD2, and obtain the fourth network parameters after the second update and the optimized two-dimensional quadratic discriminant network ; S5.5: For the one-dimensional high sea state sea clutter data set Perform random sampling again to obtain one-dimensional secondary sampled sea clutter samples , and input the one-dimensional secondary sampled sea clutter samples into the one-dimensional secondary generation network SG1 to obtain one-dimensional secondary suppression samples and the corresponding one-dimensional secondary suppression sample label set . Set the element values in the one-dimensional secondary suppression sample label set to 1; For the two-dimensional high-sea-state sea clutter dataset Perform random sampling again to obtain two-dimensional secondary sampled sea clutter samples , and input the two-dimensional secondary sampled sea clutter samples into the two-dimensional secondary generation network SG2, and output two-dimensional secondary suppression samples and the corresponding two-dimensional secondary suppression sample label set . Set the element values in the two-dimensional secondary suppression sample label set to 1; S5.6: Utilize the one-dimensional quadratic suppression samples Train the one-dimensional quadratic generation network SG1, set the third generation loss function, optimize the one-dimensional quadratic generation network SG1, and obtain the first network parameters after the second update and the optimized one-dimensional quadratic generation network ; Using the two-dimensional quadratic suppression samples Train the two-dimensional quadratic generation network SG2, set the fourth generation loss function, optimize the two-dimensional quadratic generation network SG2, and obtain the second network parameters after the second update and the optimized two-dimensional quadratic generation network ; S5.7: Utilize the optimized one-dimensional quadratic generation network , the optimized one-dimensional quadratic discriminant network , the optimized two-dimensional quadratic generation network S and the optimized two-dimensional quadratic discriminant network S to obtain a secondarily optimized high sea state sea clutter generation neural network model.

7. A method for suppressing sea clutter in high sea states based on deep learning according to claim 6, characterized in that, The specific method for training the generation network in steps S3.7 and S5.6 is as follows: Using the one-dimensional quadratic generated sample set The specific method for training the one-dimensional generation network G1 is as follows: Input the one-dimensional quadratic generated sample set into the optimized one-dimensional discriminant network to obtain the first generated prediction label . Use the first generated prediction label and the one-dimensional quadratic generated sample label set to calculate the value of the first generated loss function, optimize the one-dimensional generation network G1, and obtain the updated first network parameters and the optimized one-dimensional generation network ; Using the two-dimensional quadratic generated sample set The specific method for training the two-dimensional generation network G2 is as follows: Input the two-dimensional quadratic generated sample set into the optimized two-dimensional discriminant network to obtain the second generated prediction label . Use the second generated prediction label and the two-dimensional quadratic generated sample label set to calculate the value of the second generated loss function, optimize the two-dimensional generation network G2, and obtain the updated second network parameters and the optimized two-dimensional generation network ; Using the one-dimensional quadratic suppression sample The specific method for training the one-dimensional quadratic generation network SG1 is as follows: Input the one-dimensional quadratic suppression samples into the optimized one-dimensional quadratic discrimination network to obtain the third generated prediction label . Use the third generated prediction label and the one-dimensional quadratic suppression sample label set to calculate the value of the third loss function, optimize the one-dimensional quadratic generation network SG1, and obtain the first network parameters after the second update and the optimized one-dimensional quadratic generation network ; Using the two-dimensional quadratic suppression sample The specific method for training the two-dimensional quadratic generation network SG2 is as follows: Input the two-dimensional quadratic suppression samples into the optimized two-dimensional quadratic discriminant network to obtain the fourth generated prediction label . Use the fourth generated prediction label and the two-dimensional quadratic suppression sample label set to calculate the value of the fourth generated loss function, optimize the two-dimensional quadratic generation network SG2, and obtain the second network parameters after secondary update and the optimized two-dimensional quadratic generation network .

8. A method for suppressing sea clutter in high sea states based on deep learning according to claim 7, characterized in that, The loss functions in steps S3.5 and S3.7 are specifically: The first discriminant loss function is specifically: Among them, represents the first discriminant loss function value, and respectively represent the data in the one-dimensional real sample set and the one-dimensional noise sample set ; respectively represent the mathematical expectations of the one-dimensional real sample data and the one-dimensional noise sample data distribution functions. The second discriminant loss function is specifically: Among them, represents the value of the second discriminant loss function, and respectively represent the two-dimensional true sample set and the two-dimensional noise sample set in the data, respectively represent the mathematical expectations of the two-dimensional true sample data and the two-dimensional noise sample data distribution functions; The first generation loss function is specifically: Among them, represents the value of the first generation loss function, represents the data in the one-dimensional noise sample set of the second sampling, represents the mathematical expectation of the data distribution function of the one-dimensional noise sample data of the second sampling; The second generation loss function is specifically: Among them, represents the value of the second generation loss function, represents the data in the two-dimensional noise sample set of the second sampling, represents the mathematical expectation of the data distribution function of the two-dimensional noise sample of the second sampling.

9. A method for suppressing sea clutter in high sea states based on deep learning according to claim 8, characterized in that, The loss functions in steps S5.4 and S5.6 are specifically: The third discriminant loss function is specifically: Among them, is the data in the measured low sea state sea clutter data set in it, is the data in the one-dimensional high sea state sea clutter data set in it, respectively represent the mathematical expectations of the distribution functions of the measured low sea state sea clutter data and the one-dimensional high sea state sea clutter data; The fourth discriminant loss function is specifically: Among them, is the value of the fourth discrimination loss function, is the two-dimensional low sea state sea clutter data in is the two-dimensional high sea state sea clutter data in and respectively represent the mathematical expectations of the distribution functions of the two-dimensional low sea state sea clutter data and the two-dimensional high sea state sea clutter data; The third generation loss function is specifically: Among them, is the value of the third generation loss function, represents the dataset of the one-dimensional high sea state sea clutter data of the subsampling, represents the mathematical expectation of the distribution function of the one-dimensional high sea state sea clutter data of the subsampling; The fourth generation loss function is specifically: Among them, is the value of the fourth generation loss function, represents the dataset of the two-dimensional high sea state sea clutter data of the sub-sampling, denotes the mathematical expectation of the distribution function of the two-dimensional high sea state sea clutter data of the sub-sampling.

10. A method for suppressing sea clutter in high sea states based on deep learning according to claim 9, characterized in that, In step S6, the high sea state sea clutter data to be suppressed is obtained, and the twice-optimized high sea state sea clutter generation neural network model is used to suppress the high sea state sea clutter data to be suppressed. The specific method is as follows: S6.1: Measure and obtain the high sea state sea clutter data T0 to be suppressed; S6.2: Input the high sea clutter data T0 to be suppressed into the optimized one-dimensional quadratic generation network to obtain the one-dimensional suppressed high sea clutter data T1; S6.3: Perform time-frequency spectrum conversion on the one-dimensional suppressed high sea state sea clutter data T1 to obtain two-dimensional high sea state sea clutter data T2; S6.4: Input the two-dimensional high-sea-state sea clutter data T2 into the optimized two-dimensional quadratic generation network Perform two-dimensional suppression to obtain the suppressed high-sea-state sea clutter data.

Images